insidethemachine.ai

# Inside The Machine

> Agent-first directory of 268 local-AI tools, models and repos (runtimes, agents, memory, RAG, routing, MCP, UIs, voice, image/video, fine-tuning, quantization, evals, scaffolds, observability, hardware utilities). 13 "runbook" listings have install/verify/rollback checked against official docs; the rest are "catalog" entries with verified metadata and unverified, README-extracted install commands. Gap Scores are pending test on Mac Studio.

Append .md or .json to any /setups/<slug> URL, or send Accept: text/markdown.

## Start: the system builder (setup sheets)

The home page (https://insidethemachine.ai/) asks how you run it (fully local, or Claude conducts) and three more questions, and builds one setup sheet, A to Z: the local model; the app that runs it (the harness: the program that runs the AI and gives it tools); the phone channel, with the allow-list step first; images, voice and video on your computer, with a cloud fallback (Kie.ai, fal.ai) at the provider's listed price; the MCP servers (plug-ins that give the AI one more tool) for the job; the steps for the job; and a starter AGENTS.md. Status: every setup sheet is not tested yet. Each step says so. What we do not know says "unknown".

0. How do you want to run it? (mode)
   - `local`: Fully local. Free, private, works offline.
   - `claude`: Claude conducts. Claude Code plans and checks. Your computer does the heavy lifting. Needs a paid Claude plan. Pro is $20 a month ($17 a month if you pay for a year): https://claude.com/pricing. What leaves your computer: your prompts, and what Claude reads and writes, go to Anthropic. https://code.claude.com/docs/en/data-usage
1. What do you want to do? (pack: one or more, up to 5; several packs make one sheet, one app, one local model and one workspace folder)
   - `assistant`: Personal assistant. Reminders, notes and web look-ups. Use it in a chat window, or text it from your phone.
   - `social`: Social media growth, inside the rules. Turn work you already made into more posts, by each platform's rules.
   - `cold-email`: Cold email with Instantly, done properly. A few honest, researched business emails, sent through Instantly with your OK.
   - `code`: Coding and game dev helper. A coding helper that builds a small game with you, on your computer.
   - `creator`: Creator studio. Make images, short clips and voice-overs on your own computer.
2. How do you want to use it? (way)
   - `phone`: Text it from your phone. Message it like a friend: Telegram, WhatsApp, iMessage or Signal. In Claude mode: Control Claude Code from the Claude app on your phone (Remote Control). Telegram, Discord and iMessage channels are a research preview and may change.
   - `chat`: A chat window on your computer. Type to it in a window, like ChatGPT, but it runs on your computer.
   - `terminal`: In the terminal. Type to it in the Terminal app, inside the folder you work in.
   - `always`: Running 24/7. It works on a schedule while you sleep.
3. What machine do you have? (machine = `<os>-<gb>gb`)
   - os: `mac`, `windows`, `linux`. gb: 8, 16, 24, 32, 48, 64, 96, 128, 192, 256, 512 (default 16). On a PC with a graphics card, use the card's memory (VRAM). Not one of these sizes? Use the largest one at or below yours (36 GB counts as 32, a 12 GB graphics card as 8).
   - Not sure? The read-only scan (`curl -fsSL https://insidethemachine.ai/scan.sh | sh`, Windows: `irm https://insidethemachine.ai/scan.ps1 | iex`) prints a "Setup sheet:" link.

- Setup sheet: https://insidethemachine.ai/start/<packs joined by +>/<way>/<os>-<gb>gb[/claude] (31 sets of packs × 4 ways × 33 machines × 2 modes = 8,184 sheets, each built on its first request). Packs joined by + in the order of the packs list (another order is redirected). No /claude at the end means fully local; one pack with no /claude is the original URL. Add .md for markdown or .json for data, or send Accept: text/markdown | application/json to the sheet URL.
- Examples: https://insidethemachine.ai/start/assistant/phone/mac-16gb.md · https://insidethemachine.ai/start/code/terminal/linux-32gb.md · https://insidethemachine.ai/start/creator/chat/windows-24gb.md · https://insidethemachine.ai/start/assistant+code/terminal/mac-32gb/claude.md · https://insidethemachine.ai/start/social+cold-email+creator/chat/windows-24gb.md
- From answers: https://insidethemachine.ai/start/go?mode=<local|claude>&pack=<pack>&pack=<pack>&way=<way>&os=<os>&gb=<gb> (pack repeated or comma-separated; add &ch=whatsapp|imessage|signal for a phone channel other than Telegram on a fully local sheet) redirects to the sheet, or to the front door with the answers it has. It also takes the scan's query (ram, vram, os, chip, gpu), m=<machine id>:<GB> for a machine from https://insidethemachine.ai/machines.json (it sets the system and the size that machine counts as), and any gb (snapped down to one of our sizes).
- Starter download for a sheet: https://insidethemachine.ai/starter/<packs joined by +>-<way>-<os>-<gb>gb[-claude].md (its AGENTS.md, saved as AGENTS.md). The whole folder is a .zip: the human page offers it behind an email form (The Local Drop, unsubscribe any time). Agents: use AGENTS.md and the tree in the sheet's .json. The folder is a set of rooms, each a pipeline of numbered stages, with a CONTEXT.md for every room and every stage. Credit: Jake Van Clief's two videos (https://youtu.be/MkN-ss2Nl10, https://youtu.be/m7en7ugujlI) and the paper by Van Clief and McDermott (arXiv 2603.16021, https://arxiv.org/abs/2603.16021), in our own words. The run files, the Brain line, the gates and the session files are Inside The Machine's own.
- The front door in markdown: https://insidethemachine.ai/index.md (also / with Accept: text/markdown).

## Start here

- [Zero to working local AI](https://insidethemachine.ai/start.md): 5-step runbook + AGENTS.md/SOUL.md + lean-context guide. JSON: https://insidethemachine.ai/start.json
- [Models vs memory](https://insidethemachine.ai/models.md): which open models fit which RAM tier. JSON: https://insidethemachine.ai/models.json (130 text models; plus speech.json and images.json)
- [Machines](https://insidethemachine.ai/machines.md): hardware guide. JSON: https://insidethemachine.ai/machines.json

## Explore: what fits your computer

Page: https://insidethemachine.ai/ (the home page; https://insidethemachine.ai/explore redirects there) · markdown: https://insidethemachine.ai/explore.md (with the free shop: model and tool aisles, the My setup cart and checkout) · JSON: https://insidethemachine.ai/explore.json

The facts the home page shows in FIG. 2 (What fits), for every memory size. Default memory on the page: 16 GB.

Source: models.json 2026-10-26, speech.json 2026-10-23, images.json 2026-10-16. Scores: Artificial Analysis Intelligence Index, 2026-10-15; "% of best paid" = score / Claude Opus 5.5 (57.6).

Fit rule (model lists and the fit labels on this page): We count 75% of your memory as usable for AI. The rest is for the system and your apps. Runs well: the model fits in usable memory with at least 10% to spare. Runs, but tight: it fits, but less than 10% of usable memory is left. Close other apps first. Needs N GB+: it does not fit. N is the smallest memory size it fits in. Size unknown: we do not know how much memory it needs.
Recipe rule (the recipe sheets): Runs well: the whole recipe fits, and the model it installs leaves at least 10% of usable memory free. Runs, but tight: it fits, but the memory is nearly full. Close other apps first. Needs N GB+: it does not fit. N is the smallest memory size where it does. Each size gets the smartest model that fits. Where that model nearly fills the memory, the label says tight, even when a smaller size runs well. The sizes count the whole recipe (model and apps), and each label is the worse of the recipe rule and the model rule for the model it installs, the same on every page. (Sizes: lib/goals.ts minGb / goodGb; model rule: lib/facts.ts fitState on the model the recipe installs.)

## Our pick per memory size

The smartest model that fits and runs with Ollama on Mac, Windows and Linux. The chat setup sheets install it; the 24/7 helper below may use a smaller brain so the ears and voice fit next to it.

- 8 GB: Qwen3.5 4B, 23% of Claude Opus 5.5, needs 5.8 GB
- 16 GB: Gemma 4 12B, 25% of Claude Opus 5.5, needs 11.2 GB
- 24 GB: Gemma 4 12B, 25% of Claude Opus 5.5, needs 11.2 GB
- 32 GB: Qwen3.8 27B, 59% of Claude Opus 5.5, needs 22.7 GB
- 48 GB: Qwen3.8 27B, 59% of Claude Opus 5.5, needs 22.7 GB
- 64 GB: Qwen3.8 27B, 59% of Claude Opus 5.5, needs 22.7 GB
- 96 GB: Qwen3.8 27B, 59% of Claude Opus 5.5, needs 22.7 GB
- 128 GB: Qwen3.8 27B, 59% of Claude Opus 5.5, needs 22.7 GB
- 192 GB: Qwen3.8-Flash-Next (125B-A6B), 69% of Claude Opus 5.5, needs 140 GB
- 256 GB: Qwen3.8-Flash-Next (125B-A6B), 69% of Claude Opus 5.5, needs 140 GB
- 512 GB: Qwen3.8-Flash-Next (125B-A6B), 69% of Claude Opus 5.5, needs 140 GB

## Your 24/7 helper (FIG. 2a)

Rule: The brain, the ears (speech to text) and the voice (text to speech) stay loaded together. Ears and voice get up to 15% of the memory for AI each (at least 1 GB). The brain is the best model that can use tools, fits in the rest and runs with Ollama on Mac, Windows and Linux. To make an image or a video, the helper pauses the brain (ollama stop) and that model uses its space.

| memory | brain (always on) | ears | voice | always on | spare | image (on demand) | video (on demand) |
|---|---|---|---|---|---|---|---|
| 8 GB | Qwen3.5 2B 4.2 GB, 12% of Claude Opus 5.5, `ollama run qwen3.5:2b` | Qwen3-ASR 0.6B 1 GB | Kokoro 82M v1.0 0.5 GB | 5.7 of 6 GB | 0.3 GB | Z-Image Turbo (4 GB), runs in stable-diffusion.cpp, ComfyUI or InvokeAI. Data note: 4 GB with low-bit GGUF in stable-diffusion.cpp; maker says full model fits 16 GB. | none fits (no video model in our list fits yet) |
| 16 GB | Qwen3.5 9B 9.6 GB, 23% of Claude Opus 5.5, `ollama run qwen3.5:9b` | Qwen3-ASR 0.6B 1 GB | Kokoro 82M v1.0 0.5 GB | 11.1 of 12 GB | 0.9 GB | Z-Image Turbo (4 GB), runs in stable-diffusion.cpp, ComfyUI or InvokeAI. Data note: 4 GB with low-bit GGUF in stable-diffusion.cpp; maker says full model fits 16 GB. | none fits (no video model in our list fits yet) |
| 24 GB | Gemma 4 12B 11.2 GB, 25% of Claude Opus 5.5, `ollama run gemma4:12b` | MOSS-Transcribe-Diarize 0.9B 2.5 GB | Magpie TTS Multilingual 357M 2 GB | 15.7 of 18 GB | 2.3 GB | Z-Image Turbo (4 GB), runs in stable-diffusion.cpp, ComfyUI or InvokeAI. Data note: 4 GB with low-bit GGUF in stable-diffusion.cpp; maker says full model fits 16 GB. | none fits (no video model in our list fits yet) |
| 32 GB | Gemma 4 12B 11.2 GB, 25% of Claude Opus 5.5, `ollama run gemma4:12b` | Qwen3-ASR 1.7B 3 GB | Magpie TTS Multilingual 357M 2 GB | 16.2 of 24 GB | 7.8 GB | Z-Image Turbo (4 GB), runs in stable-diffusion.cpp, ComfyUI or InvokeAI. Data note: 4 GB with low-bit GGUF in stable-diffusion.cpp; maker says full model fits 16 GB. | HunyuanVideo 1.5 (14 GB), runs in ComfyUI or stable-diffusion.cpp. Data note: 14 GB GPU with offloading; license excludes the EU, UK and South Korea. |
| 48 GB | Qwen3.8 27B 22.7 GB, 59% of Claude Opus 5.5, `ollama run qwen3.8:27b` | Qwen3-ASR 1.7B 3 GB | Magpie TTS Multilingual 357M 2 GB | 27.7 of 36 GB | 8.3 GB | FLUX.2 [dev] (24 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 24 GB is the maker's 4-bit example on an RTX 4090 with a remote text encoder. | HunyuanVideo 1.5 (14 GB), runs in ComfyUI or stable-diffusion.cpp. Data note: 14 GB GPU with offloading; license excludes the EU, UK and South Korea. |
| 64 GB | Qwen3.8 27B 22.7 GB, 59% of Claude Opus 5.5, `ollama run qwen3.8:27b` | Granite Speech 5.0 2B 6 GB | Magpie TTS Multilingual 357M 2 GB | 30.7 of 48 GB | 17.3 GB | FLUX.2 [dev] (24 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 24 GB is the maker's 4-bit example on an RTX 4090 with a remote text encoder. | HunyuanVideo 1.5 (14 GB), runs in ComfyUI or stable-diffusion.cpp. Data note: 14 GB GPU with offloading; license excludes the EU, UK and South Korea. |
| 96 GB | Qwen3.8 27B 22.7 GB, 59% of Claude Opus 5.5, `ollama run qwen3.8:27b` | Granite Speech 5.0 2B 6 GB | Breeze TTS 2 8 GB | 36.7 of 72 GB | 35.3 GB | Qwen-Image 2.1 (46 GB), runs in stable-diffusion.cpp, ComfyUI or mflux (Mac only). Data note: Memory is mflux bf16 peak (64 GB Mac comfortable, -q 8 for less); research-only license. | MiniMax H3 (42.5 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 42.5 GB is ComfyUI's smallest variant (123.6 GB full); also tops open models on AA text-to-video at 1159. |
| 128 GB | Qwen3.8 27B 22.7 GB, 59% of Claude Opus 5.5, `ollama run qwen3.8:27b` | Granite Speech 5.0 2B 6 GB | Breeze TTS 2 8 GB | 36.7 of 96 GB | 59.3 GB | Qwen-Image 2.1 (46 GB), runs in stable-diffusion.cpp, ComfyUI or mflux (Mac only). Data note: Memory is mflux bf16 peak (64 GB Mac comfortable, -q 8 for less); research-only license. | MiniMax H3 (42.5 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 42.5 GB is ComfyUI's smallest variant (123.6 GB full); also tops open models on AA text-to-video at 1159. |
| 192 GB | Qwen3.8 27B 22.7 GB, 59% of Claude Opus 5.5, `ollama run qwen3.8:27b` | Granite Speech 5.0 2B 6 GB | Breeze TTS 2 8 GB | 36.7 of 144 GB | 107.3 GB | Qwen-Image 2.1 (46 GB), runs in stable-diffusion.cpp, ComfyUI or mflux (Mac only). Data note: Memory is mflux bf16 peak (64 GB Mac comfortable, -q 8 for less); research-only license. | MiniMax H3 (42.5 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 42.5 GB is ComfyUI's smallest variant (123.6 GB full); also tops open models on AA text-to-video at 1159. |
| 256 GB | Qwen3.8-Flash-Next (125B-A6B) 140 GB, 69% of Claude Opus 5.5, `ollama run qwen3.8-flash-next:125b-a6b` | Granite Speech 5.0 2B 6 GB | Fish Audio Drama 3 (preview) 28 GB | 174 of 192 GB | 18 GB | Qwen-Image 2.1 (46 GB), runs in stable-diffusion.cpp, ComfyUI or mflux (Mac only). Data note: Memory is mflux bf16 peak (64 GB Mac comfortable, -q 8 for less); research-only license. | MiniMax H3 (42.5 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 42.5 GB is ComfyUI's smallest variant (123.6 GB full); also tops open models on AA text-to-video at 1159. |
| 512 GB | Qwen3.8-Flash-Next (125B-A6B) 140 GB, 69% of Claude Opus 5.5, `ollama run qwen3.8-flash-next:125b-a6b` | Granite Speech 5.0 2B 6 GB | Fish Audio Drama 3 (preview) 28 GB | 174 of 384 GB | 210 GB | Qwen-Image 2.1 (46 GB), runs in stable-diffusion.cpp, ComfyUI or mflux (Mac only). Data note: Memory is mflux bf16 peak (64 GB Mac comfortable, -q 8 for less); research-only license. | MiniMax H3 (42.5 GB), runs in ComfyUI, stable-diffusion.cpp or InvokeAI. Data note: 42.5 GB is ComfyUI's smallest variant (123.6 GB full); also tops open models on AA text-to-video at 1159. |

Set it up: 1. A private ChatGPT (https://insidethemachine.ai/do/chat) installs Ollama, the brain and a chat app. 2. Speech to text (https://insidethemachine.ai/do/voice) adds the ears. It installs Whisper, a different speech-to-text model from the one on the map; both turn speech into text. 3. Voice tools (https://insidethemachine.ai/setups/c/voice) list the text-to-speech apps for the voice. We have no recipe for it yet. 4. An agent that does tasks (https://insidethemachine.ai/do/agent) adds an always-on agent app from 32 GB. On 16 GB, use the chat app from step 1. (Shown for 16 GB, the page's default size.)

- Note, 8 GB: No video model in our list fits yet. The first memory size where one fits next to the ears and voice is 32 GB.
- Note, 16 GB: No video model in our list fits yet. The first memory size where one fits next to the ears and voice is 32 GB.
- Note, 24 GB: No video model in our list fits yet. The first memory size where one fits next to the ears and voice is 32 GB.

## What do you want to do? (recipes; the page: https://insidethemachine.ai/do)

6 recipes, one sheet each, labelled for the memory you pick (default 16 GB) with the recipe rule above. Each sheet prints both sizes and how far we checked it, and has a one-command install. Labels for every size: https://insidethemachine.ai/do.md.

1. A private ChatGPT (Ollama + Open WebUI): Chat, write and brainstorm. Nothing leaves your computer. About 10 min. Needs 8 GB; runs well on 24, 48–128, 256+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/chat.sh | bash`. Tested by us: Checked: the commands were run end to end by us. Run end to end by hand on an Apple M4 Mac with 16 GB on 2026-10-15. The Open WebUI start line got --host 127.0.0.1 on 2026-10-18 (it keeps the window on this computer); that change was not re-run. https://insidethemachine.ai/do/chat.md
2. A coding helper (Ollama + OpenCode + Aider): An assistant that reads and edits your code. About 15 min. Needs 16 GB; runs well on 48–128, 256+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/code.sh | bash`. Partly tested: Partly checked: some paths were run, the rest only checked against docs. The under-32 GB path (Aider) was run end to end on an Apple M4 Mac with 16 GB on 2026-10-15. The OpenCode path is checked against the docs only. https://insidethemachine.ai/do/code.md
3. Chat with your files (Ollama + Open WebUI + Docling): Ask questions about your PDFs, notes and documents. About 20 min. Needs 16 GB; runs well on 48–128, 256+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/docs.sh | bash`. Partly tested: Partly checked: some paths were run, the rest only checked against docs. Install steps run on an Apple M4 Mac with 16 GB on 2026-10-15. File upload and the first Docling run are checked against the docs only. The Open WebUI start line got --host 127.0.0.1 on 2026-10-18 (it keeps the window on this computer); that change was not re-run. https://insidethemachine.ai/do/docs.md
4. Speech to text (whisper.cpp + Handy): Turn meetings and voice notes into text, offline. About 10 min. Needs 8 GB; runs well on 16+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/voice.sh | bash`. Tested by us: Checked: the commands were run end to end by us. Run end to end by hand on an Apple M4 Mac with 16 GB on 2026-10-15. https://insidethemachine.ai/do/voice.md
5. Make images (Draw Things + ComfyUI): Create pictures from a text prompt on your own machine. About 20 min. Needs 16 GB; runs well on 32+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/images.sh | bash`. Not tested yet: Not run: taken from the README, not checked. Point-and-click apps (Draw Things, ComfyUI). The script only checks and prints where to get them. https://insidethemachine.ai/do/images.md
6. An agent that does tasks (Hermes Agent + MCP Servers (official)): It remembers you, uses tools and learns new skills. About 30 min. Needs 32 GB; runs well on 64–128, 256+ GB. Install, then check: `curl -fsSL https://insidethemachine.ai/install/agent.sh | bash`. Checked against docs: Not run: commands checked against the official docs only. Commands checked against each tool's official docs on 2026-10-15. Not yet run end to end. https://insidethemachine.ai/do/agent.md

The scripts: The wrapper scripts are generated from the site data. Every .sh passed bash -n and shellcheck and ran in plan mode on Linux (Debian 13). Every .ps1 passed the PowerShell parser and chat.ps1 ran in plan, verify and rollback mode under PowerShell 7 on Linux with Windows calls mocked. Not run on macOS or real Windows. (bash -n and shellcheck read a script for mistakes without running it.)

## How good is AI on your own computer? (the honest gap; per memory size: https://insidethemachine.ai/best)

Good, not the best. Every model is scored on one public test, the Artificial Analysis Intelligence Index (scores read 2026-10-15). The best paid AI, Claude Opus 5.5, scores 57.6. We call that 100%. Each line: the smartest model that fits that much memory, as a share of that score (models.json 2026-10-26).

- 8 GB: Qwen3.5 4B, 23% of the best paid AI, gap 77%.
- 16–24 GB (same top pick): K2 Horizon 7B (Mac only), 36% of the best paid AI, gap 64%. Mac only: K2 Horizon 7B scores higher, but it has no Ollama download yet, only an MLX one, which needs a Mac. The recipes install Gemma 4 12B, the smartest model here that runs with Ollama on Mac, Windows and Linux. Gemma 4 12B scores 25%.
- 32–128 GB (same top pick): Qwen3.8 27B, 59% of the best paid AI, gap 41%. More memory buys longer chats and room for helpers, not a smarter brain.
- 192–256 GB (same top pick): Qwen3.8-Flash-Next (125B-A6B), 69% of the best paid AI, gap 31%.
- 512 GB: GLM-5.3-Flash (Mac only), 73% of the best paid AI, gap 27%. Mac only: GLM-5.3-Flash scores higher, but Ollama only offers it in the cloud, so it runs on your computer only with MLX, which needs a Mac. The recipes install Qwen3.8-Flash-Next (125B-A6B), the smartest model here that runs with Ollama on Mac, Windows and Linux. Qwen3.8-Flash-Next (125B-A6B) scores 69%.

Best free model: MiMo-V2.6-Pro reaches 80%. But it needs a server with 708.7 GB of memory, not a home computer.

Per memory size (the readout next to the chart):
- 8 GB: Qwen3.5 4B, 23% (gap 77%), needs 5.8 GB; 20 of 121 chat models fit. Next step up: at 16 GB the smartest model becomes K2 Horizon 7B at 36%, Mac only. The recipes install Gemma 4 12B at 25% there: it runs on Mac, Windows and Linux.
- 16 GB: K2 Horizon 7B (Mac only), 36% (gap 64%), needs 8.6 GB; 40 of 121 chat models fit. More memory, same brain: 24 GB has the same top pick. The next smarter model comes at 32 GB: Qwen3.8 27B at 59%.
- 24 GB: K2 Horizon 7B (Mac only), 36% (gap 64%), needs 8.6 GB; 47 of 121 chat models fit. Next step up: at 32 GB the smartest model becomes Qwen3.8 27B at 59%.
- 32 GB: Qwen3.8 27B, 59% (gap 41%), needs 22.7 GB; 64 of 121 chat models fit. More memory, same brain: 48 GB has the same top pick. The next smarter model comes at 192 GB: Qwen3.8-Flash-Next (125B-A6B) at 69%.
- 48 GB: Qwen3.8 27B, 59% (gap 41%), needs 22.7 GB; 80 of 121 chat models fit. More memory, same brain: 64 GB has the same top pick. The next smarter model comes at 192 GB: Qwen3.8-Flash-Next (125B-A6B) at 69%.
- 64 GB: Qwen3.8 27B, 59% (gap 41%), needs 22.7 GB; 81 of 121 chat models fit. More memory, same brain: 96 GB has the same top pick. The next smarter model comes at 192 GB: Qwen3.8-Flash-Next (125B-A6B) at 69%.
- 96 GB: Qwen3.8 27B, 59% (gap 41%), needs 22.7 GB; 87 of 121 chat models fit. More memory, same brain: 128 GB has the same top pick. The next smarter model comes at 192 GB: Qwen3.8-Flash-Next (125B-A6B) at 69%.
- 128 GB: Qwen3.8 27B, 59% (gap 41%), needs 22.7 GB; 94 of 121 chat models fit. Next step up: at 192 GB the smartest model becomes Qwen3.8-Flash-Next (125B-A6B) at 69%.
- 192 GB: Qwen3.8-Flash-Next (125B-A6B), 69% (gap 31%), needs 140 GB; 98 of 121 chat models fit. More memory, same brain: 256 GB has the same top pick. The next smarter model comes at 512 GB: GLM-5.3-Flash at 73%, Mac only. The recipes still install Qwen3.8-Flash-Next (125B-A6B) there: it runs on Mac, Windows and Linux.
- 256 GB: Qwen3.8-Flash-Next (125B-A6B), 69% (gap 31%), needs 140 GB; 99 of 121 chat models fit. Next step up: at 512 GB the smartest model becomes GLM-5.3-Flash at 73%, Mac only. The recipes still install Qwen3.8-Flash-Next (125B-A6B) there: it runs on Mac, Windows and Linux.
- 512 GB: GLM-5.3-Flash (Mac only), 73% (gap 27%), needs 233 GB; 107 of 121 chat models fit. Top of the range: 512 GB is the biggest memory size in our data.

- Good for: Everyday chat, writing, summaries, simple code, transcription.
- Still behind on: Hard reasoning, big coding jobs, long multi-step tasks.
- Why run it anyway: Private, works offline, no monthly bill, no usage limits.

Next: we test each recipe on a 128 GB Mac Studio and publish a Gap Score for it: how close the local setup gets to the best paid AI. No scores published yet: every catalog.json gap_score.value is null (268 entries, "pending test on Mac Studio"). Method: https://insidethemachine.ai/gap-score.md

Go deeper: 130 open text models, plus 73 speech and image models (https://insidethemachine.ai/models.md) · 56 machines compared (https://insidethemachine.ai/machines.md) · 36 apps that drive a model (https://insidethemachine.ai/harnesses.md) · all 268 tools (https://insidethemachine.ai/setups) · new on Hugging Face, not checked yet (https://insidethemachine.ai/my) · advanced agent guide (https://insidethemachine.ai/guides/agent.md)

## New in our data this week (Rev. 2026-10-26; the page: https://insidethemachine.ai/changes)

What changed in our data files in week 2026-W44, dated and sourced (changes.json 2026-10-26). "Fits in" is the smallest memory size whose usable part (75%) holds the model.

- New model: Nemotron 3.5 Bolt E4B. Needs 11.1 GB; fits in 16 GB (2026-10-26) https://insidethemachine.ai/models?tier=16&find=nemotron-3.5-bolt-e4b#all
- New model: Nanbeige4.3 33B. Needs 25.9 GB; fits in 48 GB (2026-10-26) https://insidethemachine.ai/models?tier=48&find=nanbeige4.3-33b#all
- Tool data changed: Ollama (2026-10-26) https://insidethemachine.ai/setups/ollama
- Model data changed: DeepSeek V4.1 Flash 1017. Needs 528.6 GB: too big for our sizes (2026-10-26) https://insidethemachine.ai/models?find=deepseek-v4.1-flash-1017#all
- Tool data changed: llama.cpp (2026-10-26) https://insidethemachine.ai/setups/llama-cpp

5 of 5 changes shown, new items first. A tool that changed in two of our files is listed once. From our weekly news check, merged into the data on 26 Oct 2026. 5 changes in week 2026-W44, 26 Oct 2026 to 26 Oct 2026 (5 log entries). All changes: https://insidethemachine.ai/changes.md

## The Local Drop

What is new in local AI, every week. New models, tools, harnesses and hardware, and what each one needs to run. From Kit, the AI agent behind our YouTube channel. Free, by email or on YouTube.

1. New models: Open models released this week, and how much memory each one needs.
2. Tools and harnesses: New apps and agents that run those models on your own computer.
3. Hardware: New machines and graphics cards, and which models fit on them.
4. Gap Scores: From November: how close each local setup gets to the best paid AI, tested on a 128 GB Mac Studio.

Gap Score status: no scores published yet. Each one is waiting for its test on the Mac Studio. Sign up: https://insidethemachine.ai/#report

Every /best, /compare, /alternatives and /machines/<id> page also has a .json twin (the same data as its .md) and answers Accept: application/json.

## Best picks (answers with dated data)

- [Best local LLM by memory](https://insidethemachine.ai/best.md): one table, 8 to 512 GB: models that fit, best model, best coding model
- [Best local LLM for 8 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-8gb-ram.md): top picks, coding/agent/vision pick, machines with 8 GB
- [Best local LLM for 16 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-16gb-ram.md): top picks, coding/agent/vision pick, machines with 16 GB
- [Best local LLM for 24 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-24gb-ram.md): top picks, coding/agent/vision pick, machines with 24 GB
- [Best local LLM for 32 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-32gb-ram.md): top picks, coding/agent/vision pick, machines with 32 GB
- [Best local LLM for 48 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-48gb-ram.md): top picks, coding/agent/vision pick, machines with 48 GB
- [Best local LLM for 64 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-64gb-ram.md): top picks, coding/agent/vision pick, machines with 64 GB
- [Best local LLM for 96 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-96gb-ram.md): top picks, coding/agent/vision pick, machines with 96 GB
- [Best local LLM for 128 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-128gb-ram.md): top picks, coding/agent/vision pick, machines with 128 GB
- [Best local LLM for 192 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-192gb-ram.md): top picks, coding/agent/vision pick, machines with 192 GB
- [Best local LLM for 256 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-256gb-ram.md): top picks, coding/agent/vision pick, machines with 256 GB
- [Best local LLM for 512 GB RAM or VRAM](https://insidethemachine.ai/best/local-llm-for-512gb-ram.md): top picks, coding/agent/vision pick, machines with 512 GB
- [Best local coding model](https://insidethemachine.ai/best/local-coding-model.md): per memory size, Terminal-Bench, coding apps
- [Best local model for agents](https://insidethemachine.ai/best/local-model-for-agents.md): OpenClaw, Hermes Agent and more
- [Best local LLM for Mac](https://insidethemachine.ai/best/local-llm-for-mac.md) · [NVIDIA GPU](https://insidethemachine.ai/best/local-llm-for-nvidia-gpu.md) · [mini PC](https://insidethemachine.ai/best/local-llm-for-mini-pc.md)
- [Compare apps](https://insidethemachine.ai/compare.md): [Ollama vs LM Studio](https://insidethemachine.ai/compare/ollama-vs-lm-studio.md) · [Ollama vs llama.cpp](https://insidethemachine.ai/compare/ollama-vs-llama-cpp.md) · [LM Studio vs llama.cpp](https://insidethemachine.ai/compare/lm-studio-vs-llama-cpp.md) · [MLX LM vs Ollama](https://insidethemachine.ai/compare/mlx-lm-vs-ollama.md) · [Open WebUI vs LM Studio](https://insidethemachine.ai/compare/open-webui-vs-lm-studio.md)
- Alternatives: [Ollama](https://insidethemachine.ai/alternatives/ollama.md) · [MLX LM](https://insidethemachine.ai/alternatives/mlx-lm.md) · [LM Studio](https://insidethemachine.ai/alternatives/lm-studio.md) · [llama.cpp](https://insidethemachine.ai/alternatives/llama-cpp.md) · [Aider](https://insidethemachine.ai/alternatives/aider.md) · [Hermes Agent](https://insidethemachine.ai/alternatives/hermes-agent.md) · [LiteLLM](https://insidethemachine.ai/alternatives/litellm.md) · [Mem0](https://insidethemachine.ai/alternatives/mem0.md) · [Docling](https://insidethemachine.ai/alternatives/docling.md) · [Open WebUI](https://insidethemachine.ai/alternatives/open-webui.md) · [whisper.cpp](https://insidethemachine.ai/alternatives/whisper-cpp.md)
- Machines (one page each): [MacBook Air 13"/15" (M5)](https://insidethemachine.ai/machines/macbook-air-m5.md) · [MacBook Pro 14" (M5)](https://insidethemachine.ai/machines/macbook-pro-14-m5.md) · [MacBook Pro 14"/16" (M5 Pro)](https://insidethemachine.ai/machines/macbook-pro-m5-pro.md) · [MacBook Pro 14"/16" (M5 Max)](https://insidethemachine.ai/machines/macbook-pro-m5-max.md) · [Mac mini (M6)](https://insidethemachine.ai/machines/mac-mini-m6.md) · [Mac mini (M5 Pro)](https://insidethemachine.ai/machines/mac-mini-m5-pro.md) · [Mac Studio (M5 Max)](https://insidethemachine.ai/machines/mac-studio-m5-max.md) · [Mac Studio (M5 Ultra)](https://insidethemachine.ai/machines/mac-studio-m5-ultra.md) · [GeForce RTX 5090](https://insidethemachine.ai/machines/rtx-5090.md) · [GeForce RTX 5080](https://insidethemachine.ai/machines/rtx-5080.md) · [GeForce RTX 5070 Ti](https://insidethemachine.ai/machines/rtx-5070-ti.md) · [GeForce RTX 5070](https://insidethemachine.ai/machines/rtx-5070.md) · [GeForce RTX 5060 Ti 16GB](https://insidethemachine.ai/machines/rtx-5060-ti-16gb.md) · [GeForce RTX 5060](https://insidethemachine.ai/machines/rtx-5060.md) · [GeForce RTX 4090](https://insidethemachine.ai/machines/rtx-4090.md) · [GeForce RTX 4080 Super](https://insidethemachine.ai/machines/rtx-4080-super.md) · [GeForce RTX 4070 Ti Super](https://insidethemachine.ai/machines/rtx-4070-ti-super.md) · [GeForce RTX 4060 Ti 16GB](https://insidethemachine.ai/machines/rtx-4060-ti-16gb.md) · [GeForce RTX 3090 (used)](https://insidethemachine.ai/machines/rtx-3090.md) · [GeForce RTX 3090 Ti](https://insidethemachine.ai/machines/rtx-3090-ti.md) · [GeForce RTX 3060 12GB (used)](https://insidethemachine.ai/machines/rtx-3060-12gb.md) · [RTX PRO 6000 Blackwell Workstation (96GB)](https://insidethemachine.ai/machines/rtx-pro-6000-blackwell.md) · [RTX PRO 6000 Blackwell Max-Q (96GB)](https://insidethemachine.ai/machines/rtx-pro-6000-blackwell-maxq.md) · [RTX PRO 5000 Blackwell 72GB](https://insidethemachine.ai/machines/rtx-pro-5000-blackwell-72gb.md) · [RTX PRO 4500 Blackwell 32GB](https://insidethemachine.ai/machines/rtx-pro-4500-blackwell.md) · [RTX PRO 4000 Blackwell 24GB](https://insidethemachine.ai/machines/rtx-pro-4000-blackwell.md) · [NVIDIA DGX Spark (128GB)](https://insidethemachine.ai/machines/dgx-spark-128gb.md) · [DGX Spark 64GB (partner GB10 systems)](https://insidethemachine.ai/machines/dgx-spark-64gb.md) · [NVIDIA Jetson AGX Thor Developer Kit](https://insidethemachine.ai/machines/jetson-agx-thor.md) · [NVIDIA Jetson Orin Nano Super Developer Kit](https://insidethemachine.ai/machines/jetson-orin-nano-super.md) · [Framework Desktop (Ryzen AI Max+ 395)](https://insidethemachine.ai/machines/framework-desktop-395.md) · [GMKtec EVO-X2 (Ryzen AI Max+ 395)](https://insidethemachine.ai/machines/gmktec-evo-x2.md) · [Minisforum MS-S1 MAX (Ryzen AI Max+ 395)](https://insidethemachine.ai/machines/minisforum-ms-s1-max.md) · [Bosgame M5 (Ryzen AI Max+ 395)](https://insidethemachine.ai/machines/bosgame-m5.md) · [Beelink GTR9 Pro (Ryzen AI Max+ 395)](https://insidethemachine.ai/machines/beelink-gtr9-pro.md) · [GMKtec EVO-X5 Pro (Ryzen AI Max+ PRO 495 "Gorgon Halo")](https://insidethemachine.ai/machines/gmktec-evo-x5-pro.md) · [Radeon RX 7900 XTX](https://insidethemachine.ai/machines/radeon-rx-7900-xtx.md) · [Radeon RX 9070 XT](https://insidethemachine.ai/machines/radeon-rx-9070-xt.md) · [Radeon AI PRO R9700 (32GB)](https://insidethemachine.ai/machines/radeon-ai-pro-r9700.md) · [Intel Arc B580](https://insidethemachine.ai/machines/intel-arc-b580.md) · [Intel Arc Pro B60 (24GB)](https://insidethemachine.ai/machines/intel-arc-pro-b60.md) · [Intel Arc Pro B70 (32GB)](https://insidethemachine.ai/machines/intel-arc-pro-b70.md) · [Raspberry Pi 5 (16GB), CPU only](https://insidethemachine.ai/machines/raspberry-pi-5-16gb.md) · [Tesla P40 24GB (used)](https://insidethemachine.ai/machines/tesla-p40.md) · [AMD Instinct MI50 32GB (used)](https://insidethemachine.ai/machines/instinct-mi50-32gb.md) · [Tesla V100 PCIe 32GB (used)](https://insidethemachine.ai/machines/tesla-v100-32gb.md) · [Minisforum MS-S1 MAX-P495 (Ryzen AI Max+ PRO 495 "Gorgon Halo")](https://insidethemachine.ai/machines/minisforum-ms-s1-max-p495.md) · [Gigabyte W775-V10-L01 (DGX Station GB300)](https://insidethemachine.ai/machines/gigabyte-w775-v10-l01.md) · [ASUS ProArt GR1X (NVIDIA RTX Spark)](https://insidethemachine.ai/machines/asus-proart-gr1x.md) · [iBUYPOWER Pro AMD AI Halo Mini PC (Ryzen AI Max+ PRO 495)](https://insidethemachine.ai/machines/ibuypower-pro-amd-ai-halo.md) · [Microsoft Surface RTX Spark Dev Box (128GB)](https://insidethemachine.ai/machines/surface-rtx-spark-dev-box.md) · [AMD Ryzen AI Halo Developer Platform (Ryzen AI Max+ PRO 495, 192GB)](https://insidethemachine.ai/machines/amd-ryzen-ai-halo-495.md) · [Sipeed MaixStation](https://insidethemachine.ai/machines/sipeed-maixstation.md) · [NVIDIA DGX Station GB300 (partner systems)](https://insidethemachine.ai/machines/nvidia-dgx-station-gb300.md) · [Dell XPS 16 Creator Edition (NVIDIA RTX Spark)](https://insidethemachine.ai/machines/dell-xps-16-creator-edition.md)

## Your machine, one-command installs, what changed

- [Scan my machine](https://insidethemachine.ai/scan.sh): read-only; prints a https://insidethemachine.ai/my?... link and a setup sheet link (https://insidethemachine.ai/start/go?...). Windows: https://insidethemachine.ai/scan.ps1. Report for agents: https://insidethemachine.ai/my.json?ram=<GB>&os=<mac|linux|win>&chip=<slug>&gpu=<slug>&vram=<GB> (also /my.md)
- [One-command installs](https://insidethemachine.ai/install.md): 19 scripts with install, verify, rollback and an honest run status. JSON: https://insidethemachine.ai/install.json
- [Install: A private ChatGPT](https://insidethemachine.ai/install/chat.md) [ran]: `curl -fsSL https://insidethemachine.ai/install/chat.sh | bash`
- [Install: A coding helper](https://insidethemachine.ai/install/code.md) [partly-ran]: `curl -fsSL https://insidethemachine.ai/install/code.sh | bash`
- [Install: Chat with your files](https://insidethemachine.ai/install/docs.md) [partly-ran]: `curl -fsSL https://insidethemachine.ai/install/docs.sh | bash`
- [Install: Speech to text](https://insidethemachine.ai/install/voice.md) [ran]: `curl -fsSL https://insidethemachine.ai/install/voice.sh | bash`
- [Install: Make images](https://insidethemachine.ai/install/images.md) [not-run]: `curl -fsSL https://insidethemachine.ai/install/images.sh | bash`
- [Install: An agent that does tasks](https://insidethemachine.ai/install/agent.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/agent.sh | bash`
- [Install: MLX LM](https://insidethemachine.ai/install/mlx-lm.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/mlx-lm.sh | bash`
- [Install: Ollama](https://insidethemachine.ai/install/ollama.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/ollama.sh | bash`
- [Install: llama.cpp](https://insidethemachine.ai/install/llama-cpp.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/llama-cpp.sh | bash`
- [Install: LM Studio](https://insidethemachine.ai/install/lm-studio.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/lm-studio.sh | bash`
- [Install: Open WebUI](https://insidethemachine.ai/install/open-webui.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/open-webui.sh | bash`
- [Install: LiteLLM](https://insidethemachine.ai/install/litellm.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/litellm.sh | bash`
- [Install: Aider](https://insidethemachine.ai/install/aider.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/aider.sh | bash`
- [Install: Hermes Agent](https://insidethemachine.ai/install/hermes-agent.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/hermes-agent.sh | bash`
- [Install: Mem0](https://insidethemachine.ai/install/mem0.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/mem0.sh | bash`
- [Install: MCP Servers (official)](https://insidethemachine.ai/install/mcp-servers.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/mcp-servers.sh | bash`
- [Install: whisper.cpp](https://insidethemachine.ai/install/whisper-cpp.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/whisper-cpp.sh | bash`
- [Install: Docling](https://insidethemachine.ai/install/docling.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/docling.sh | bash`
- [Install: AGENTS.md](https://insidethemachine.ai/install/agents-md.md) [docs-checked]: `curl -fsSL https://insidethemachine.ai/install/agents-md.sh | bash`
- The status in brackets is how far we ran each recipe's or runbook's steps ourselves, not the script: [ran] Checked: the commands were run end to end by us; [partly-ran] Partly checked: some paths were run, the rest only checked against docs; [docs-checked] Not run: commands checked against the official docs only; [not-run] Not run: taken from the README, not checked.
- How far the scripts themselves were checked (install.json wrapper): The wrapper scripts are generated from the site data. Every .sh passed bash -n and shellcheck and ran in plan mode on Linux (Debian 13). Every .ps1 passed the PowerShell parser and chat.ps1 ran in plan, verify and rollback mode under PowerShell 7 on Linux with Windows calls mocked. Not run on macOS or real Windows. (bash -n and shellcheck read a script for mistakes without running it.)
- Every script takes three forms: as is (install, then check); with `-s -- plan` after `bash` (prints every step, changes nothing); with `-s -- rollback --yes` after `bash` (removes what it added).
- [What changed](https://insidethemachine.ai/changes.md): dated new models (with the memory they need), tools, machines, and what broke. JSON: https://insidethemachine.ai/changes.json · RSS: https://insidethemachine.ai/changes.xml

## Pages (every page also answers Accept: text/markdown)

- [Home: build your own local AI helper](https://insidethemachine.ai/index.md): the one front door: what fits your memory, then three ways on.
- [The free shop](https://insidethemachine.ai/shop.md): model and tool aisles, sorted by what fits; add parts to My setup, then check out. JSON: https://insidethemachine.ai/shop.json
- [Skills for your agent](https://insidethemachine.ai/skills.md): skill packs ranked by GitHub stars gained, with install lines. JSON: https://insidethemachine.ai/skills.json
- [Checkout](https://insidethemachine.ai/checkout.md): how you run it (fully local, or Claude conducts), then three questions (one or more packs, one or more ways, the machine; the cart answers some), then one setup sheet, or one combined setup for several ways; one email gives the folder and the prompt. JSON: https://insidethemachine.ai/index.json (also / with Accept: application/json)
- [Setup sheets](https://insidethemachine.ai/start/assistant/phone/mac-16gb.md): one A-to-Z setup per set of packs (one to five), mode (local or Claude conducts), way and machine; the URL pattern is in "Start" above. Not tested yet. JSON: add .json
- [Learn: local AI in plain words](https://insidethemachine.ai/learn.md): the guides (The basics, the advanced agent guide; the Gap Score method is coming). JSON: https://insidethemachine.ai/learn.json
- [Community](https://insidethemachine.ai/community.md): the Local Build community, where people will launch the local AI tools they build (opens November 2026). JSON: https://insidethemachine.ai/community.json
- [Explore: the free shop](https://insidethemachine.ai/explore.md): what fits per memory size; model and tool aisles; the My setup cart and checkout; links to recipes, the gap, what is new and the directory. JSON: https://insidethemachine.ai/explore.json
- [Start here: local AI in plain words](https://insidethemachine.ai/start-here.md): primer + recipe list + the full runbook. JSON: https://insidethemachine.ai/start-here.json
- [Recipes](https://insidethemachine.ai/do.md): 6 goals with memory needs. JSON: https://insidethemachine.ai/do.json
- [Recipe: A private ChatGPT](https://insidethemachine.ai/do/chat.md): needs 8 GB+, best with 16 GB+, about 10 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/chat.json
- [Recipe: A coding helper](https://insidethemachine.ai/do/code.md): needs 16 GB+, best with 32 GB+, about 15 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/code.json
- [Recipe: Chat with your files](https://insidethemachine.ai/do/docs.md): needs 16 GB+, best with 32 GB+, about 20 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/docs.json
- [Recipe: Speech to text](https://insidethemachine.ai/do/voice.md): needs 8 GB+, best with 16 GB+, about 10 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/voice.json
- [Recipe: Make images](https://insidethemachine.ai/do/images.md): needs 16 GB+, best with 32 GB+, about 20 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/images.json
- [Recipe: An agent that does tasks](https://insidethemachine.ai/do/agent.md): needs 32 GB+, best with 64 GB+, about 30 minutes. Steps, model per memory size, rollback, fixes. JSON: https://insidethemachine.ai/do/agent.json
- [Harnesses](https://insidethemachine.ai/harnesses.md): which app for which job, Claude Code vs Codex with local models. JSON: https://insidethemachine.ai/harnesses.json
- [Advanced: an agent that improves itself](https://insidethemachine.ai/guides/agent.md): same as start.md. JSON: https://insidethemachine.ai/guides/agent.json
- [Gap Score](https://insidethemachine.ai/gap-score.md): methodology, no scores yet. JSON: https://insidethemachine.ai/gap-score.json

## Build: business recipes for solo creators

Claude conducts, your machine builds. Claude does the planning, the hard writing and the final review. That takes few tokens. Local models do the bulk work on every item, for free and in private. Every recipe shows receipts: Claude tokens (this recipe vs doing it all in Claude), local minutes and dollars.

- [Build index](https://insidethemachine.ai/build.md): families, blueprints, tools, receipts
- [Blueprint: Shorts cutter](https://insidethemachine.ai/build/shorts-cutter.md): Turn one long video into a handful of Shorts. Your machine does the listening and cutting. Claude picks the best moments. Not run yet. JSON: https://insidethemachine.ai/build/shorts-cutter.json
- [Blueprint: Videos → newsletter](https://insidethemachine.ai/build/videos-to-newsletter.md): Turn this week's videos into a newsletter draft. Your machine drafts each section. Claude writes the intro and does the final edit. Not run yet. JSON: https://insidethemachine.ai/build/videos-to-newsletter.json
- [Blueprint: Thumbnail & title lab](https://insidethemachine.ai/build/thumbnail-title-lab.md): Make many title and thumbnail ideas on your machine, score them all, and let Claude pick the few worth testing. Not run yet. JSON: https://insidethemachine.ai/build/thumbnail-title-lab.json
- [Blueprint: Sponsor inbox](https://insidethemachine.ai/build/sponsor-inbox.md): Sort every sponsor email on your own machine. Real offers rise to the top, scams and mass pitches drop out. Nothing leaves your computer. Not run yet. JSON: https://insidethemachine.ai/build/sponsor-inbox.json
- [Tool: Local Jev](https://insidethemachine.ai/build/local-jev.md): A decide-only model on your own machine. It answers typed questions about each item: pick one, give a score, or yes/no. It never writes. Runs today (CPU test, speed not measured). JSON: https://insidethemachine.ai/build/local-jev.json
- [Tool: Claude's local intern](https://insidethemachine.ai/build/local-intern.md): A small MCP server that lets Claude Code hand bulk jobs to your own machine: draft, sort, transcribe, caption. Each job comes back with a token receipt. Runs today (CPU test, speed not measured). JSON: https://insidethemachine.ai/build/local-intern.json
- [Privacy](https://insidethemachine.ai/privacy.md): what Share your run sends and never sends

## Site MCP server

https://insidethemachine.ai/mcp (remote MCP over Streamable HTTP, no key needed to read).

- Claude Code: `claude mcp add --transport http insidethemachine https://insidethemachine.ai/mcp`
- Cursor, Windsurf and other clients that take a URL: `{"mcpServers":{"insidethemachine":{"url":"https://insidethemachine.ai/mcp"}}}`
- Clients that only run local commands (stdio): `npx -y mcp-remote https://insidethemachine.ai/mcp`

Tools:
- `what_fits`: What runs on a memory size or a machine: the smartest model, the model for each recipe part, and which build recipes run there.
- `search`: Search recipes, models, machines and tools.
- `list_recipes`: Every build recipe, with its status.
- `get_recipe`: One recipe: Claude's job, the local parts per memory size, the steps and the receipts.
- `get_model`: One model: memory need, score and run command.
- `get_runs`: Approved runs for a recipe. Only runs a person has approved.
- `submit_run`: Share your run: numbers only (machine, models, stage times, token counts). It goes to a review queue first.

## Core

- [Catalog JSON](https://insidethemachine.ai/api/index.json): every listing with links
- [Listing JSON Schema](https://insidethemachine.ai/schema/listing.schema.json)
- [Full text](https://insidethemachine.ai/llms-full.txt)
- [Dropped / archived items](https://insidethemachine.ai/api/dropped.json)
- [About / Gap Score methodology](https://insidethemachine.ai/about.md)

## The tools directory

Part numbers (T-001 to T-268) follow the stack: layer (model, runtime, harness, memory, tools), then group, then A to Z by name. Star counts never move a number; a new tool shifts the numbers after it. Labels in brackets: [Apple Silicon] Built for Macs with M-series chips. [Cloud] Uses cloud models: sends your prompts to an AI in the cloud. [Closed source] You cannot read its code. [runbook] = install, verify and rollback checked against official docs.

## Runtimes & inference

- [Ollama](https://insidethemachine.ai/setups/ollama.md): T-054 · One command to download and run a model. [runbook]
- [SGLang](https://insidethemachine.ai/setups/sglang.md): T-057 · High-throughput serving for LLMs and VLMs.
- [MLX](https://insidethemachine.ai/setups/mlx.md): T-049 · Apple's array framework for Apple silicon. [Apple Silicon]
- [llamafile](https://insidethemachine.ai/setups/llamafile.md): T-044 · Package a model and runtime into one executable.
- [MLC LLM](https://insidethemachine.ai/setups/mlc-llm.md): T-048 · Compile and run LLMs on any device.
- [KoboldCpp](https://insidethemachine.ai/setups/koboldcpp.md): T-041 · One-file GGUF runner with a built-in UI.
- [MLX Examples](https://insidethemachine.ai/setups/mlx-examples.md): T-050 · Official MLX example models and scripts. [Apple Silicon]
- [MLX LM](https://insidethemachine.ai/setups/mlx-lm.md): T-051 · Runs LLMs natively on Apple silicon. [runbook] [Apple Silicon]
- [llama-swap](https://insidethemachine.ai/setups/llama-swap.md): T-042 · Hot-swaps models behind one OpenAI-style endpoint.
- [MLX-VLM](https://insidethemachine.ai/setups/mlx-vlm.md): T-053 · Vision-language models on MLX. [Apple Silicon]
- [LM Studio](https://insidethemachine.ai/setups/lm-studio.md): T-045 · A desktop app for local models. [runbook]
- [RamaLama](https://insidethemachine.ai/setups/ramalama.md): T-056 · Serve models from any source in containers.
- [vLLM Metal](https://insidethemachine.ai/setups/vllm-metal.md): T-059 · vLLM plugin for Apple Silicon.
- [MLX OpenAI Server](https://insidethemachine.ai/setups/mlx-openai-server.md): T-052 · OpenAI-compatible API server for MLX models. [Apple Silicon]
- [ML Drift](https://insidethemachine.ai/setups/ml-drift.md): T-047 · Google's GPU inference engine for edge devices. [Apple Silicon]
- [llama.cpp](https://insidethemachine.ai/setups/llama-cpp.md): T-043 · The engine behind most local AI. [runbook]
- [vLLM](https://insidethemachine.ai/setups/vllm.md): T-058 · High-throughput LLM serving engine.
- [PyTorch](https://insidethemachine.ai/setups/pytorch.md): T-055 · The main toolkit AI apps are built on; runs models on Mac GPUs. [Apple Silicon]
- [LocalAI](https://insidethemachine.ai/setups/localai.md): T-046 · Runs text, voice, image and video models behind one OpenAI-style address. [Apple Silicon]

## Model hubs

- [Transformers](https://insidethemachine.ai/setups/transformers.md): T-007 · The reference library for open model weights.
- [Unsloth notebooks](https://insidethemachine.ai/setups/unsloth-notebooks.md): T-009 · 250+ ready fine-tuning notebooks.
- [Hugging Face Hub](https://insidethemachine.ai/setups/hugging-face-hub.md): T-001 · The main hub for open model weights.
- [Ollama Library](https://insidethemachine.ai/setups/ollama-library.md): T-006 · Ready-to-pull models for Ollama.
- [LM Studio model catalog](https://insidethemachine.ai/setups/lm-studio-models.md): T-003 · Curated models you can load in LM Studio.
- [mlx-community (Hugging Face)](https://insidethemachine.ai/setups/mlx-community.md): T-005 · Community MLX conversions of popular models. [Apple Silicon]
- [Unsloth (Hugging Face)](https://insidethemachine.ai/setups/unsloth-models.md): T-008 · Unsloth GGUF and fine-tune-ready model uploads.
- [lmstudio-community (Hugging Face)](https://insidethemachine.ai/setups/lmstudio-community.md): T-004 · LM Studio's community GGUF/MLX uploads.
- [Kaggle Models](https://insidethemachine.ai/setups/kaggle-models.md): T-002 · Google-run model hub with downloadable weights.

## Models from the channel

- [Qwen-Image-2.1-Turbo](https://insidethemachine.ai/setups/qwen-image-2-1-turbo.md): T-015 · 7B image model that paints and edits in 8 steps.
- [Mellum2.1 12B-A2.5B](https://insidethemachine.ai/setups/mellum-2-1.md): T-014 · JetBrains coding model, 12B with 2.5B active.
- [EmbeddingGemma 2](https://insidethemachine.ai/setups/embeddinggemma-2.md): T-012 · 740M on-device embedding model for search.
- [Kolibri-1](https://insidethemachine.ai/setups/kolibri-1.md): T-013 · Open German/English MoE model (about 78 GB of memory, per the video).
- [AstaBrief 8B](https://insidethemachine.ai/setups/astabrief-8b.md): T-010 · Small Ai2 model that writes cited reports.
- [Clef](https://insidethemachine.ai/setups/cloudflare-clef.md): T-011 · Cloudflare's small open decision models (Clef, Clef-flash).

## Coding agents

- [OpenCode](https://insidethemachine.ai/setups/opencode.md): T-090 · Open-source coding agent for your terminal.
- [Pi (pi-mono)](https://insidethemachine.ai/setups/pi-pi-mono.md): T-092 · Small agent toolkit: LLM API, agent loop and coding CLI.
- [Goose](https://insidethemachine.ai/setups/goose.md): T-087 · Extensible open-source agent that runs code and tools.
- [Aider](https://insidethemachine.ai/setups/aider.md): T-079 · A coding agent in your terminal. [runbook]
- [mini-SWE-agent](https://insidethemachine.ai/setups/mini-swe-agent.md): T-089 · A 100-line agent that fixes GitHub issues.
- [Claude Code](https://insidethemachine.ai/setups/claude-code.md): T-080 · Anthropic's terminal coding agent (cloud model). [Cloud]
- [Gemini CLI](https://insidethemachine.ai/setups/gemini-cli.md): T-085 · Google's open-source terminal agent for Gemini. [Cloud]
- [OpenHands](https://insidethemachine.ai/setups/openhands.md): T-091 · Open platform for autonomous software agents.
- [Cline](https://insidethemachine.ai/setups/cline.md): T-081 · Coding agent as a VS Code extension, CLI or SDK.
- [Continue](https://insidethemachine.ai/setups/continue.md): T-083 · Open-source coding agent for IDEs and CI.
- [Tabby](https://insidethemachine.ai/setups/tabby.md): T-094 · Self-hosted code completion and chat server.
- [Crush](https://insidethemachine.ai/setups/crush.md): T-084 · Polished terminal coding agent from Charm.
- [Qwen Code](https://insidethemachine.ai/setups/qwen-code.md): T-093 · Qwen's open-source terminal coding agent.
- [Kilo Code](https://insidethemachine.ai/setups/kilo-code.md): T-088 · All-in-one agentic coding extension and CLI.
- [Codex CLI](https://insidethemachine.ai/setups/codex-cli.md): T-082 · OpenAI's lightweight terminal coding agent.
- [GitHub Copilot CLI](https://insidethemachine.ai/setups/copilot-cli.md): T-086 · GitHub's terminal coding agent; can use Ollama models. [Cloud]

## General & personal agents

- [OpenClaw](https://insidethemachine.ai/setups/openclaw.md): T-106 · Personal agent that runs on your machine and chats anywhere.
- [Hermes Agent](https://insidethemachine.ai/setups/hermes-agent.md): T-100 · A personal agent that learns skills and remembers you. [runbook]
- [AutoGPT](https://insidethemachine.ai/setups/autogpt.md): T-095 · Platform for building autonomous agents.
- [Browser Use](https://insidethemachine.ai/setups/browser-use.md): T-096 · Let an AI agent drive a web browser.
- [DeerFlow](https://insidethemachine.ai/setups/deerflow.md): T-098 · Long-running research and coding super-agent.
- [Open Interpreter](https://insidethemachine.ai/setups/open-interpreter.md): T-105 · Agent that runs code on your computer, for open models.
- [OpenManus](https://insidethemachine.ai/setups/openmanus.md): T-107 · Open-source general agent (Manus-style).
- [nanobot](https://insidethemachine.ai/setups/nanobot.md): T-103 · Ultra-light self-hosted personal agent.
- [UI-TARS Desktop](https://insidethemachine.ai/setups/ui-tars-desktop.md): T-109 · Multimodal agent that operates your desktop.
- [Khoj](https://insidethemachine.ai/setups/khoj.md): T-101 · Self-hostable AI second brain over your docs and the web.
- [Perplexica](https://insidethemachine.ai/setups/perplexica.md): T-108 · Open answering engine (formerly Perplexica).
- [NanoClaw](https://insidethemachine.ai/setups/nanoclaw.md): T-104 · Lightweight, container-sandboxed OpenClaw alternative.
- [GPT Researcher](https://insidethemachine.ai/setups/gpt-researcher.md): T-099 · Agent that writes deep research reports.
- [Cua](https://insidethemachine.ai/setups/cua.md): T-097 · Open-source computer-use drivers and sandboxes.
- [LM Studio Bionic](https://insidethemachine.ai/setups/lm-studio-bionic.md): T-102 · LM Studio's agent that can run beside the app. [Closed source] [Apple Silicon]

## Agent frameworks

- [LiteLLM](https://insidethemachine.ai/setups/litellm.md): T-118 · One API in front of every model, local or cloud. [runbook]
- [AutoGen](https://insidethemachine.ai/setups/autogen.md): T-111 · Microsoft's multi-agent programming framework.
- [CrewAI](https://insidethemachine.ai/setups/crewai.md): T-112 · Role-based multi-agent framework.
- [LangGraph](https://insidethemachine.ai/setups/langgraph.md): T-116 · Build stateful, resilient agent graphs.
- [Agno](https://insidethemachine.ai/setups/agno.md): T-110 · Build and run agent systems in Python.
- [DSPy](https://insidethemachine.ai/setups/dspy.md): T-113 · Program LLMs instead of hand-writing prompts.
- [OpenAI Agents SDK](https://insidethemachine.ai/setups/openai-agents-sdk.md): T-120 · OpenAI's lightweight multi-agent SDK.
- [smolagents](https://insidethemachine.ai/setups/smolagents.md): T-122 · Tiny library for agents that think in code.
- [Mastra](https://insidethemachine.ai/setups/mastra.md): T-119 · TypeScript framework for agents and workflows.
- [Vercel AI SDK](https://insidethemachine.ai/setups/vercel-ai-sdk.md): T-123 · TypeScript toolkit for AI apps and agents.
- [Haystack](https://insidethemachine.ai/setups/haystack.md): T-115 · Orchestration framework for RAG and agents.
- [Letta](https://insidethemachine.ai/setups/letta.md): T-117 · Stateful agents with long-term memory (ex-MemGPT).
- [Google ADK](https://insidethemachine.ai/setups/google-adk.md): T-114 · Google's code-first agent development kit.
- [Pydantic AI](https://insidethemachine.ai/setups/pydantic-ai.md): T-121 · Typed Python agent framework.

## Memory

- [Mem0](https://insidethemachine.ai/setups/mem0.md): T-182 · Long-term memory for your agent. [runbook]
- [claude-mem](https://insidethemachine.ai/setups/claude-mem.md): T-177 · Captures agent sessions and replays context later.
- [Cognee](https://insidethemachine.ai/setups/cognee.md): T-178 · Knowledge-graph memory for agents.
- [Graphiti](https://insidethemachine.ai/setups/graphiti.md): T-179 · Real-time temporal knowledge graphs for agents.
- [Supermemory](https://insidethemachine.ai/setups/supermemory.md): T-186 · Fast memory engine and app, can run locally.
- [MemOS](https://insidethemachine.ai/setups/memos.md): T-184 · Memory operating system for LLM agents.
- [Honcho](https://insidethemachine.ai/setups/honcho.md): T-180 · Memory library that models users across sessions.
- [Basic Memory](https://insidethemachine.ai/setups/basic-memory.md): T-176 · Markdown-file memory your agent reads and writes.
- [Memobase](https://insidethemachine.ai/setups/memobase.md): T-183 · User-profile memory for chat apps.
- [memsearch](https://insidethemachine.ai/setups/memsearch.md): T-185 · Semantic search over agent memory files.
- [MCP Memory Service](https://insidethemachine.ai/setups/mcp-memory-service.md): T-181 · Persistent memory exposed as an MCP server.

## RAG & document parsing

- [Docling](https://insidethemachine.ai/setups/docling.md): T-189 · Turns documents into clean text for AI. [runbook]
- [Marker](https://insidethemachine.ai/setups/marker.md): T-192 · Fast, accurate PDF-to-markdown converter.
- [LightRAG](https://insidethemachine.ai/setups/lightrag.md): T-191 · Simple, fast graph-based RAG.
- [Qdrant](https://insidethemachine.ai/setups/qdrant.md): T-198 · High-performance vector database.
- [Chroma](https://insidethemachine.ai/setups/chroma.md): T-188 · Open-source embedding database.
- [pgvector](https://insidethemachine.ai/setups/pgvector.md): T-197 · Vector search inside Postgres.
- [LanceDB](https://insidethemachine.ai/setups/lancedb.md): T-190 · Embedded multimodal vector database.
- [sqlite-vec](https://insidethemachine.ai/setups/sqlite-vec.md): T-200 · Vector search as a tiny SQLite extension.
- [MarkItDown](https://insidethemachine.ai/setups/markitdown.md): T-193 · Convert files and Office docs to markdown.
- [RAGFlow](https://insidethemachine.ai/setups/ragflow.md): T-199 · Deep-document RAG engine with a UI.
- [PaddleOCR](https://insidethemachine.ai/setups/paddleocr.md): T-196 · OCR and document parsing toolkit.
- [MinerU](https://insidethemachine.ai/setups/mineru.md): T-195 · Converts complex PDFs to markdown or JSON.
- [AnythingLLM](https://insidethemachine.ai/setups/anythingllm.md): T-187 · All-in-one desktop app for chat with your docs.
- [Meilisearch](https://insidethemachine.ai/setups/meilisearch.md): T-194 · Fast search engine with hybrid vector search.

## Routing & gateways

- [Strands Decider](https://insidethemachine.ai/setups/strands-decider.md): T-069 · Small decision model that routes agent choices.
- [New API](https://insidethemachine.ai/setups/new-api.md): T-063 · Unified gateway that converts many LLM APIs.
- [One API](https://insidethemachine.ai/setups/one-api.md): T-066 · LLM API management and distribution gateway.
- [Portkey Gateway](https://insidethemachine.ai/setups/portkey-gateway.md): T-068 · Fast AI gateway with guardrails and fallbacks.
- [Bifrost](https://insidethemachine.ai/setups/bifrost.md): T-060 · High-performance LLM and MCP gateway.
- [Plano](https://insidethemachine.ai/setups/plano.md): T-067 · AI-native proxy with smart LLM routing.
- [LLMRouter](https://insidethemachine.ai/setups/llmrouter.md): T-062 · Research library of LLM routing methods.
- [Envoy AI Gateway](https://insidethemachine.ai/setups/envoy-ai-gateway.md): T-061 · Envoy-based gateway for generative AI services.
- [NVIDIA LLM Router](https://insidethemachine.ai/setups/nvidia-llm-router.md): T-064 · Routes each prompt to the best-fit model.
- [NVIDIA PAIR](https://insidethemachine.ai/setups/nvidia-pair.md): T-065 · Shares AI requests across the computers on your home network.

## MCP servers & tools

- [MCP Servers (official)](https://insidethemachine.ai/setups/mcp-servers.md): T-210 · Standard plug-in tools for agents. [runbook]
- [MCP Python SDK](https://insidethemachine.ai/setups/mcp-python-sdk.md): T-209 · Official Python SDK for MCP servers and clients.
- [LLM CLI](https://insidethemachine.ai/setups/llm-cli.md): T-207 · CLI and Python library for any LLM, local or cloud.
- [MCP Inspector](https://insidethemachine.ai/setups/mcp-inspector.md): T-208 · Visual tester for MCP servers.
- [n8n](https://insidethemachine.ai/setups/n8n.md): T-211 · Workflow automation with AI nodes, self-hostable.
- [Firecrawl](https://insidethemachine.ai/setups/firecrawl.md): T-205 · Turn websites into LLM-ready data.
- [Awesome MCP Servers](https://insidethemachine.ai/setups/awesome-mcp-servers.md): T-201 · Big curated list of MCP servers.
- [Crawl4AI](https://insidethemachine.ai/setups/crawl4ai.md): T-204 · Open-source crawler that outputs clean markdown.
- [Context7](https://insidethemachine.ai/setups/context7.md): T-203 · Up-to-date library docs for coding agents.
- [Chrome DevTools MCP](https://insidethemachine.ai/setups/chrome-devtools-mcp.md): T-202 · Chrome DevTools as an MCP server.
- [SearXNG](https://insidethemachine.ai/setups/searxng.md): T-213 · Self-hosted private metasearch engine.
- [Playwright MCP](https://insidethemachine.ai/setups/playwright-mcp.md): T-212 · Browser automation MCP server.
- [GitHub MCP Server](https://insidethemachine.ai/setups/github-mcp-server.md): T-206 · GitHub's official MCP server.
- [Serena](https://insidethemachine.ai/setups/serena.md): T-214 · Semantic code retrieval and editing over MCP.

## Chat UIs & front-ends

- [Open WebUI](https://insidethemachine.ai/setups/open-webui.md): T-134 · A ChatGPT-style app for your local models. [runbook]
- [NextChat](https://insidethemachine.ai/setups/nextchat.md): T-133 · Lightweight cross-platform chat client.
- [LobeChat](https://insidethemachine.ai/setups/lobechat.md): T-132 · Agent workspace and chat UI.
- [Cherry Studio](https://insidethemachine.ai/setups/cherry-studio.md): T-126 · Desktop AI studio for many providers.
- [text-generation-webui](https://insidethemachine.ai/setups/text-generation-webui.md): T-137 · Desktop app for local LLMs (text-generation-webui).
- [LibreChat](https://insidethemachine.ai/setups/librechat.md): T-130 · Self-hosted multi-provider ChatGPT clone.
- [Jan](https://insidethemachine.ai/setups/jan.md): T-129 · Offline ChatGPT alternative for your desktop.
- [Chatbox](https://insidethemachine.ai/setups/chatbox.md): T-125 · Desktop and mobile AI client.
- [SillyTavern](https://insidethemachine.ai/setups/sillytavern.md): T-136 · Power-user chat and roleplay front-end.
- [Open-LLM-VTuber](https://insidethemachine.ai/setups/open-llm-vtuber.md): T-135 · Voice chat with an animated avatar, locally.
- [HF Chat UI](https://insidethemachine.ai/setups/hf-chat-ui.md): T-128 · The open code behind HuggingChat.
- [big-AGI](https://insidethemachine.ai/setups/big-agi.md): T-124 · Advanced multi-model chat suite.
- [Enchanted](https://insidethemachine.ai/setups/enchanted.md): T-127 · Native macOS/iOS app for self-hosted models. [Apple Silicon]
- [llama.vscode](https://insidethemachine.ai/setups/llama-vscode.md): T-131 · Local code completion in VS Code via llama.cpp.

## Voice: speech-to-text & TTS

- [whisper.cpp](https://insidethemachine.ai/setups/whisper-cpp.md): T-226 · Speech-to-text that runs on your machine. [runbook]
- [MLX-Audio](https://insidethemachine.ai/setups/mlx-audio.md): T-222 · TTS, STT and speech-to-speech on MLX. [Apple Silicon]
- [WhisperKit](https://insidethemachine.ai/setups/whisperkit.md): T-227 · On-device speech AI for Apple Silicon (WhisperKit). [Apple Silicon]
- [Parakeet MLX](https://insidethemachine.ai/setups/parakeet-mlx.md): T-224 · NVIDIA Parakeet speech recognition on MLX. [Apple Silicon]
- [Whisper](https://insidethemachine.ai/setups/whisper.md): T-225 · OpenAI's original speech recognition model.
- [ChatTTS](https://insidethemachine.ai/setups/chattts.md): T-216 · Conversational text-to-speech model.
- [Handy](https://insidethemachine.ai/setups/handy.md): T-221 · Offline push-to-talk dictation app.
- [Fish Speech](https://insidethemachine.ai/setups/fish-speech.md): T-220 · Open-source text-to-speech with voice cloning.
- [Chatterbox](https://insidethemachine.ai/setups/chatterbox.md): T-215 · Open-source expressive TTS.
- [faster-whisper](https://insidethemachine.ai/setups/faster-whisper.md): T-219 · Whisper transcription, faster, via CTranslate2.
- [WhisperX](https://insidethemachine.ai/setups/whisperx.md): T-228 · Whisper with word timestamps and speaker labels.
- [CosyVoice](https://insidethemachine.ai/setups/cosyvoice.md): T-217 · Multilingual voice generation model.
- [Dia](https://insidethemachine.ai/setups/dia.md): T-218 · TTS that generates realistic dialogue.
- [NeMo (Parakeet)](https://insidethemachine.ai/setups/nemo-parakeet.md): T-223 · NVIDIA NeMo speech models (incl. Parakeet).

## Image & video generation

- [stable-diffusion.cpp](https://insidethemachine.ai/setups/stable-diffusion-cpp.md): T-240 · Image and video diffusion in pure C/C++.
- [mflux](https://insidethemachine.ai/setups/mflux.md): T-236 · MLX-native image and video models for Mac.
- [Qwen-Image 2.1](https://insidethemachine.ai/setups/qwen-image-2-1.md): T-238 · Qwen's open image generation and editing model.
- [Draw Things](https://insidethemachine.ai/setups/draw-things.md): T-233 · Community repo for the Draw Things Mac/iOS app. [Apple Silicon]
- [Stable Diffusion WebUI](https://insidethemachine.ai/setups/stable-diffusion-webui.md): T-239 · The classic Stable Diffusion web UI.
- [ComfyUI](https://insidethemachine.ai/setups/comfyui.md): T-229 · Node-based GUI and backend for diffusion models.
- [Fooocus](https://insidethemachine.ai/setups/fooocus.md): T-234 · Prompt-first, low-setup image generator.
- [Diffusers](https://insidethemachine.ai/setups/diffusers.md): T-232 · PyTorch library for image, video and audio diffusion.
- [Open-Sora](https://insidethemachine.ai/setups/open-sora.md): T-237 · Open video generation models.
- [InvokeAI](https://insidethemachine.ai/setups/invokeai.md): T-235 · Professional creative studio for diffusion.
- [Core ML Stable Diffusion](https://insidethemachine.ai/setups/core-ml-stable-diffusion.md): T-231 · Stable Diffusion on Core ML for Apple Silicon. [Apple Silicon]
- [Wan 2.2](https://insidethemachine.ai/setups/wan-2-2.md): T-242 · Open video generation models (Wan 2.2).
- [Wan 2.1](https://insidethemachine.ai/setups/wan-2-1.md): T-241 · Open video generation models (Wan 2.1).
- [ComfyUI Manager](https://insidethemachine.ai/setups/comfyui-manager.md): T-230 · Install and manage ComfyUI custom nodes.

## Fine-tuning

- [MLX-LM-LoRA](https://insidethemachine.ai/setups/mlx-lm-lora.md): T-019 · LoRA and full fine-tuning on MLX. [Apple Silicon]
- [Unsloth](https://insidethemachine.ai/setups/unsloth.md): T-028 · Fast, low-memory fine-tuning with a local UI.
- [LLaMA-Factory](https://insidethemachine.ai/setups/llama-factory.md): T-017 · Fine-tune 100+ LLMs and VLMs from one tool.
- [nanoGPT](https://insidethemachine.ai/setups/nanogpt.md): T-022 · The simplest repo to train small GPTs.
- [nanochat](https://insidethemachine.ai/setups/nanochat.md): T-021 · Train a small ChatGPT end to end on a budget.
- [Open R1](https://insidethemachine.ai/setups/open-r1.md): T-023 · Open reproduction of DeepSeek-R1 training.
- [verl](https://insidethemachine.ai/setups/verl.md): T-029 · Reinforcement-learning post-training framework.
- [PEFT](https://insidethemachine.ai/setups/peft.md): T-026 · LoRA and other parameter-efficient fine-tuning.
- [TRL](https://insidethemachine.ai/setups/trl.md): T-027 · Train models with SFT, DPO and RL.
- [ms-swift](https://insidethemachine.ai/setups/ms-swift.md): T-020 · Fine-tune 600+ LLMs with PEFT or full training.
- [Axolotl](https://insidethemachine.ai/setups/axolotl.md): T-016 · Config-driven fine-tuning toolkit.
- [OpenRLHF](https://insidethemachine.ai/setups/openrlhf.md): T-024 · Scalable RLHF and agentic RL on Ray.
- [Oumi](https://insidethemachine.ai/setups/oumi.md): T-025 · Fine-tune, evaluate and deploy open models.
- [MergeKit](https://insidethemachine.ai/setups/mergekit.md): T-018 · Merge model weights into new models.

## Quantization

- [bitsandbytes](https://insidethemachine.ai/setups/bitsandbytes.md): T-032 · 4-bit and 8-bit quantization for PyTorch.
- [Core ML Tools](https://insidethemachine.ai/setups/core-ml-tools.md): T-033 · Convert and compress models for Core ML. [Apple Silicon]
- [TensorRT Model Optimizer](https://insidethemachine.ai/setups/tensorrt-model-optimizer.md): T-039 · NVIDIA quantization, pruning and distillation.
- [LLM Compressor](https://insidethemachine.ai/setups/llm-compressor.md): T-036 · Compress models for vLLM serving.
- [Optimum](https://insidethemachine.ai/setups/optimum.md): T-037 · Hardware-optimized inference and quantization for HF models.
- [torchao](https://insidethemachine.ai/setups/torchao.md): T-040 · PyTorch-native quantization.
- [ExLlamaV3](https://insidethemachine.ai/setups/exllamav3.md): T-034 · Fast EXL3 quantization and inference on consumer GPUs.
- [AutoRound](https://insidethemachine.ai/setups/autoround.md): T-031 · Accurate low-bit post-training quantization.
- [AQLM](https://insidethemachine.ai/setups/aqlm.md): T-030 · Extreme compression with additive quantization.
- [GPTQModel](https://insidethemachine.ai/setups/gptqmodel.md): T-035 · GPTQ quantization toolkit across GPU vendors.
- [Optimum Quanto](https://insidethemachine.ai/setups/optimum-quanto.md): T-038 · Simple PyTorch quantization backend.

## Evals & benchmarks

- [ThinkingBox](https://insidethemachine.ai/setups/thinkingbox.md): T-256 · Runs agents against MCP tools and checks what they really did.
- [promptfoo](https://insidethemachine.ai/setups/promptfoo.md): T-251 · Test and red-team prompts, agents and RAG.
- [OpenAI Evals](https://insidethemachine.ai/setups/openai-evals.md): T-249 · Framework and registry for evaluating LLMs.
- [DeepEval](https://insidethemachine.ai/setups/deepeval.md): T-243 · Unit-test style evaluation for LLM apps.
- [Ragas](https://insidethemachine.ai/setups/ragas.md): T-252 · Metrics for evaluating RAG pipelines.
- [LM Evaluation Harness](https://insidethemachine.ai/setups/lm-evaluation-harness.md): T-247 · The standard harness for academic LLM benchmarks.
- [Gorilla / BFCL](https://insidethemachine.ai/setups/gorilla-bfcl.md): T-244 · Function-calling models and the Berkeley tool-use leaderboard.
- [OpenCompass](https://insidethemachine.ai/setups/opencompass.md): T-250 · Large evaluation platform across many benchmarks.
- [SWE-bench](https://insidethemachine.ai/setups/swe-bench.md): T-254 · Benchmark of real GitHub issues for coding agents.
- [simple-evals](https://insidethemachine.ai/setups/simple-evals.md): T-253 · OpenAI's lightweight reference evals.
- [LMMs-Eval](https://insidethemachine.ai/setups/lmms-eval.md): T-248 · Evaluation toolkit for multimodal models.
- [Inspect AI](https://insidethemachine.ai/setups/inspect-ai.md): T-245 · UK AISI framework for LLM and agent evals.
- [Terminal-Bench](https://insidethemachine.ai/setups/terminal-bench.md): T-255 · Benchmark of hard tasks in a real terminal.
- [LightEval](https://insidethemachine.ai/setups/lighteval.md): T-246 · Hugging Face's multi-backend eval toolkit.

## Scaffolds & templates

- [AGENTS.md](https://insidethemachine.ai/setups/agents-md.md): T-138 · A README written for AI agents. [runbook]
- [ClawHub](https://insidethemachine.ai/setups/clawhub.md): T-143 · Skill and plugin registry for OpenClaw.
- [Spec Kit](https://insidethemachine.ai/setups/spec-kit.md): T-146 · Spec-driven development toolkit for agents.
- [OpenSpec](https://insidethemachine.ai/setups/openspec.md): T-144 · Lightweight spec-driven workflow for coding agents.
- [Awesome Claude Code](https://insidethemachine.ai/setups/awesome-claude-code.md): T-139 · Curated Claude Code commands, hooks and workflows.
- [BMAD Method](https://insidethemachine.ai/setups/bmad-method.md): T-141 · Agile multi-agent development method.
- [Awesome Cursor Rules](https://insidethemachine.ai/setups/awesome-cursor-rules.md): T-140 · Collection of agent rule files.
- [wshobson agents](https://insidethemachine.ai/setups/wshobson-agents.md): T-148 · Multi-harness agent plugin marketplace.
- [Claude Code Templates](https://insidethemachine.ai/setups/claude-code-templates.md): T-142 · CLI to install agent configs and templates.
- [Task Master](https://insidethemachine.ai/setups/task-master.md): T-147 · Task-management system for coding agents.
- [SOUL.md (persona file)](https://insidethemachine.ai/setups/soul-md.md): T-145 · A file that sets your agent's identity and voice.

## Skills

- [Superpowers](https://insidethemachine.ai/setups/superpowers.md): T-173 · Makes coding agents plan, test and review their work.
- [Anthropic Skills](https://insidethemachine.ai/setups/anthropic-skills.md): T-149 · Make PDFs, Word, Excel and slide files; write new skills.
- [Vercel Agent Skills](https://insidethemachine.ai/setups/vercel-agent-skills.md): T-175 · React, Next.js and web design rules from Vercel.
- [Hugging Face Skills](https://insidethemachine.ai/setups/huggingface-skills.md): T-160 · Find, run, train and share models on Hugging Face.
- [OpenAI Skills](https://insidethemachine.ai/setups/openai-skills.md): T-165 · OpenAI's skill catalog, made for Codex.
- [Scientific Agent Skills](https://insidethemachine.ai/setups/scientific-agent-skills.md): T-169 · 177 science skills: biology, chemistry, medicine.
- [OpenScience](https://insidethemachine.ai/setups/openscience.md): T-166 · A science workbench with hundreds of research skills.
- [Matt Pocock's Skills](https://insidethemachine.ai/setups/mattpocock-skills.md): T-162 · Skills for real engineering: specs, TDD, code review.
- [gstack](https://insidethemachine.ai/setups/gstack.md): T-157 · Garry Tan's Claude Code setup: CEO, designer, QA roles.
- [Trail of Bits Skills](https://insidethemachine.ai/setups/trailofbits-skills.md): T-174 · Security skills: find bugs and audit code.
- [NVIDIA Skills](https://insidethemachine.ai/setups/nvidia-skills.md): T-164 · NVIDIA skills: CUDA, robotics, simulation, RAG.
- [Microsoft Skills](https://insidethemachine.ai/setups/microsoft-skills.md): T-163 · Skills for Azure and Microsoft SDKs.
- [Gemini Skills](https://insidethemachine.ai/setups/gemini-skills.md): T-156 · Skills for building with Google's Gemini API.
- [Cloudflare Skills](https://insidethemachine.ai/setups/cloudflare-skills.md): T-152 · Skills for building on Cloudflare.
- [Supabase Skills](https://insidethemachine.ai/setups/supabase-skills.md): T-172 · Skills for Supabase and Postgres.
- [Expo Skills](https://insidethemachine.ai/setups/expo-skills.md): T-154 · Skills for Expo and React Native apps.
- [Sentry Skills](https://insidethemachine.ai/setups/sentry-skills.md): T-171 · The Sentry team's skills: reviews, commits, bug hunts.
- [Remotion Skills](https://insidethemachine.ai/setups/remotion-skills.md): T-167 · Skills for making videos in code with Remotion.
- [Callstack Skills](https://insidethemachine.ai/setups/callstack-skills.md): T-151 · React Native skills from Callstack.
- [HyperFrames](https://insidethemachine.ai/setups/hyperframes.md): T-161 · Write HTML animation and render it to an MP4 video.
- [Higgsfield Skills](https://insidethemachine.ai/setups/higgsfield-skills.md): T-159 · Images and video with GPT Image, Seedance, Kling, Veo.
- [fal Skills](https://insidethemachine.ai/setups/fal-skills.md): T-155 · Pick, prompt and chain fal.ai image and video models.
- [Black Forest Labs Skills](https://insidethemachine.ai/setups/bfl-skills.md): T-150 · FLUX image and video prompts from the FLUX makers.
- [Replicate Skills](https://insidethemachine.ai/setups/replicate-skills.md): T-168 · Find, compare and prompt image and video models.
- [ElevenLabs Skills](https://insidethemachine.ai/setups/elevenlabs-skills.md): T-153 · Voice, music and sound effects with ElevenLabs.
- [HeyGen Skills](https://insidethemachine.ai/setups/heygen-skills.md): T-158 · Talking avatar videos and dubbing with HeyGen.
- [Seedance 2.0 Skills](https://insidethemachine.ai/setups/seedance-skills.md): T-170 · Seedance 2.0 video prompts: camera, light, sound, fixes.

## Observability

- [Langfuse](https://insidethemachine.ai/setups/langfuse.md): T-261 · Open-source tracing and evals for LLM apps.
- [MLflow](https://insidethemachine.ai/setups/mlflow.md): T-263 · ML and LLM lifecycle platform with tracing.
- [Opik](https://insidethemachine.ai/setups/opik.md): T-266 · Trace, evaluate and monitor LLM apps.
- [Phoenix](https://insidethemachine.ai/setups/phoenix.md): T-267 · Open-source LLM tracing and evaluation.
- [Evidently](https://insidethemachine.ai/setups/evidently.md): T-259 · Evaluate and monitor ML and LLM systems.
- [OpenLLMetry](https://insidethemachine.ai/setups/openllmetry.md): T-265 · OpenTelemetry instrumentation for LLM apps.
- [Helicone](https://insidethemachine.ai/setups/helicone.md): T-260 · LLM observability via a one-line proxy.
- [AgentOps](https://insidethemachine.ai/setups/agentops.md): T-258 · Monitoring and cost tracking for agents.
- [Agenta](https://insidethemachine.ai/setups/agenta.md): T-257 · Prompt management, evals and observability.
- [Logfire](https://insidethemachine.ai/setups/logfire.md): T-262 · Observability for LLM and agent systems.
- [OpenLIT](https://insidethemachine.ai/setups/openlit.md): T-264 · OpenTelemetry-native tracing and guardrails.
- [Weave](https://insidethemachine.ai/setups/weave.md): T-268 · W&B toolkit for tracing and evaluating AI apps.

## Hardware & system utilities

- [Apple container](https://insidethemachine.ai/setups/apple-container.md): T-070 · Run Linux containers natively on a Mac. [Apple Silicon]
- [Stats](https://insidethemachine.ai/setups/stats.md): T-078 · macOS menu-bar system monitor. [Apple Silicon]
- [llmfit](https://insidethemachine.ai/setups/llmfit.md): T-073 · One command to see which models fit your hardware.
- [nvtop](https://insidethemachine.ai/setups/nvtop.md): T-077 · GPU monitor for Apple, AMD, Intel and NVIDIA.
- [nvitop](https://insidethemachine.ai/setups/nvitop.md): T-076 · Interactive NVIDIA GPU process viewer.
- [GPUStack](https://insidethemachine.ai/setups/gpustack.md): T-072 · GPU cluster manager for model serving.
- [NVIDIA Container Toolkit](https://insidethemachine.ai/setups/nvidia-container-toolkit.md): T-075 · Use NVIDIA GPUs inside containers.
- [mactop](https://insidethemachine.ai/setups/mactop.md): T-074 · Apple Silicon CPU/GPU/power monitor in the terminal. [Apple Silicon]
- [gguf-parser](https://insidethemachine.ai/setups/gguf-parser.md): T-071 · Inspect GGUF files and estimate memory use.

## Optional

- [YouTube: Inside The Machine](https://www.youtube.com/@InsideTheMachineSI)