# Inside The Machine: build your own local AI helper URL: https://insidethemachine.ai/ Pick free AI parts that fit your computer. Get one folder and one prompt that set them up. ## What runs where (FIG. 1) What runs where. Local AI needs no cloud: every part is already inside your computer. 1. Chip: Does the math. A faster chip gives faster answers. 2. Memory: Holds the model while it runs. It decides what fits. 3. Model: The brain. One big file you download for free. 4. App: The window you talk to, like a private ChatGPT. ## What fits on your computer (FIG. 2) 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. ## Then pick your parts (three ways on) - The free shop: 152 AI models and 264 open tools, sorted by what fits. https://insidethemachine.ai/shop (markdown: https://insidethemachine.ai/shop.md) - Your setup sheet: three questions, then one sheet with every step. https://insidethemachine.ai/checkout - Skills for your agent, by stars gained: https://insidethemachine.ai/skills (markdown: https://insidethemachine.ai/skills.md) ## The weekly email: The Local Drop The Local Drop · free, weekly. Every new AI you can run at home. One email a week. New models, tools and hardware, and what each one needs to run. Sign up on this page: https://insidethemachine.ai/#report. Gave your email at checkout? You are already on the list. --- ## Checkout (https://insidethemachine.ai/checkout) The same checkout as a full page (the My setup panel asks the same things in short), in full: URL: https://insidethemachine.ai/checkout The home page (https://insidethemachine.ai/) is the free shop: check what fits your memory, add parts to "My setup", then check out here. The cart ticks its packs and picks its brain; you answer the rest. Answer 3 questions. Get your setup sheet, step by step: the local model, the app, the phone channel, images and voice (local or cloud), the tools (MCP servers), the steps for your job and a starter AGENTS.md. On the sheet, one email gives you the folder (with MY-CART.md for your cart) and the prompt. Status: every step is not tested yet, until we run it. ## How you run it (mode: a switch above question 1) - **Fully local** (`local`, the default): Free, private, works offline. - **Claude conducts** (`claude`): 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 (read 2026-10-18). What leaves your computer: your prompts, and what Claude reads and writes, go to Anthropic. https://code.claude.com/docs/en/data-usage ## Question 1 of 3: What do you want to do? (pack: pick one or more, up to 5) - **Personal assistant**: Personal assistant (`assistant`). Reminders, notes and web look-ups. Use it in a chat window, or text it from your phone. - **Social growth**: Social media growth, inside the rules (`social`). Turn work you already made into more posts, by each platform's rules. - **Cold email (Instantly)**: Cold email with Instantly, done properly (`cold-email`). A few honest, researched business emails, sent through Instantly with your OK. (Not in the page's list of choices; a link with pack=cold-email still opens it.) - **Coding + game dev**: Coding and game dev helper (`code`). A coding helper that builds a small game with you, on your computer. - **Creator studio**: Creator studio (`creator`). Make images, short clips and voice-overs on your own computer. Several packs make one sheet, one app, one local model and one workspace folder; each pack adds its rooms. ## Question 2 of 3: How do you want to use it? (way: pick one or more) - **Text it from your phone** (`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. (https://code.claude.com/docs/en/remote-control) - **A chat window on your computer** (`chat`): Type to it in a window, like ChatGPT, but it runs on your computer. - **In the terminal** (`terminal`): Type to it in the Terminal app, inside the folder you work in. - **Running 24/7** (`always`): It works on a schedule while you sleep. - Several ways make one combined setup: one folder with one SETUP.md for all of them, and a link to each way's own sheet. URL: the way ids joined by "+", in this order: phone, chat, terminal, always (for example https://insidethemachine.ai/start/assistant/phone+terminal/mac-16gb). - Phone only, fully local: which chat app? Telegram (`telegram`, the easiest, the default), WhatsApp (`whatsapp`), iMessage (`imessage`, needs a Mac), Signal (`signal`). The sheet has all of them; add `?ch=` to its URL to open that app's steps first on the page (the .md and .json list every app). A Claude sheet has no chat-app choice. ## Question 3 of 3: What machine do you have? - System: `mac` (Mac), `windows` (Windows PC), `linux` (Linux PC). - Memory (RAM), with the URL id: 8 GB (`8`), 16 GB (`16`), 24 GB (`24`), 32 GB (`32`), 48 GB (`48`), 64 GB (`64`), 96 GB (`96`), 128 GB (`128`), 192 GB (`192`), 256 GB (`256`), 512 GB (`512`). Default 16 GB. Gaming PC? Use the graphics 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? Pick your computer from our machine list (machines.json): it sets the system and the memory size it counts as. Older Mac, or a laptop without a graphics card? Use the memory buttons above. Or look it up: Mac: Apple menu → About This Mac → Memory. Windows: Settings → System → About → Installed RAM. Graphics card: Task Manager → Performance → GPU → Dedicated GPU memory. Linux: run free -h and read the total on the Mem line. - Or scan it: `curl -fsSL https://insidethemachine.ai/scan.sh | sh` (Mac, Linux) or `irm https://insidethemachine.ai/scan.ps1 | iex` (Windows PowerShell). Read-only: it reads your memory, chip and graphics card, and prints a link to your setup sheet (https://insidethemachine.ai/start/go?…). Nothing is uploaded. Read the script first: https://insidethemachine.ai/my ## Your parts Under question 3 the page shows our picks for your answers. Change any part, or just click Build. Three parts: - **A. Brain** (the AI model): our pick first, then up to three other models that meet the same rule for the app (tool calling, context window, an Ollama download, fits your usable memory), smartest first. A model brought from the Explore shop's "My setup" cart (https://insidethemachine.ai/explore) is added as a choice when it meets the same rule. Each shows its score as a share of the best paid AI and the memory it needs. Each line about a model is built only from models.json (its maker, whether it reads pictures, whether it is smaller than our pick). - **B. App** (the helper you talk to): Hermes Agent is our pick for every pack and every way. The other choice is the app the packs used before. OpenCode is shown as "Coding specialist": Which one codes better? We have not measured it yet. We test both on the same tasks from 10 Nov 2026. In Claude conducts mode the app is Claude Code, with no other choice. - **C. Tools** (what it can do for you): "Already in " lists what the app has built in, so it is never offered as an add-on. "Unclear" means the app's own pages disagree; we do not pick a side. "Add to it" lists the tools (MCP servers) the app can add for your packs, all ticked. Untick one and its step leaves the sheet. On phone and 24/7 sheets, the app's own tools that its lock step takes away from the bot are not listed as built in; where the bot needs one (files, memory), the matching server is an add-on. The choices as JSON: https://insidethemachine.ai/api/parts?packs=&mode=&way=&os=&gb=[&app=][&brain=]. Parts change the sheet URL only when they differ from our picks: https://insidethemachine.ai/start///-gb[/claude]/with///. A part that is not a choice for that sheet falls back to our pick. Example (Coding + game dev, terminal, Mac 32 GB): our picks are Qwen3.8 27B in Hermes Agent. With Qwen3.6 27B in OpenCode and one add-on: https://insidethemachine.ai/start/code/terminal/mac-32gb/with/opencode/qwen3.6-27b/git ## Your setup sheet URL: https://insidethemachine.ai/start///-gb[/claude] (add .md or .json for agents). Packs go in the order of the list above (assistant, social, cold-email, code, creator); another order or a repeat is redirected to the right URL. No `/claude` at the end means fully local; a single pack with no `/claude` is the original URL. 31 sets of packs × 4 ways × 33 machines (3 systems × 11 memory sizes) × 2 modes = 8,184 sheets. Examples: https://insidethemachine.ai/start/assistant/phone/mac-16gb (one pack, local) · https://insidethemachine.ai/start/assistant+code/terminal/mac-32gb/claude (two packs, Claude conducts) · https://insidethemachine.ai/start/social+cold-email+creator/chat/windows-24gb (three packs, local) From answers: https://insidethemachine.ai/start/go?mode=&pack=&pack=&way=&os=&gb= (pack repeated or comma-separated) redirects to the sheet. ## Everything else The free shop (130 open text models, 36 apps, 56 machines, 268 tools, 6 recipes): https://insidethemachine.ai/ (https://insidethemachine.ai/index.md). Business recipes for solo creators: https://insidethemachine.ai/build. New to local AI? The basic words in 5 minutes: https://insidethemachine.ai/start-here.md. The full site index for agents: https://insidethemachine.ai/llms.txt ## Get The Local Drop (the signup box on the home page) Free weekly email: new local AI models, tools, harnesses and hardware, and what each one needs to run. Unsubscribe any time. Tick box: "Also invite me to the Local Build community when it opens in November 2026." Humans sign up on the page; there is no signup API for agents. Privacy: https://insidethemachine.ai/privacy.md --- That is the whole home page. The shop: https://insidethemachine.ai/shop.md. Skills: https://insidethemachine.ai/skills.md. The questions alone: https://insidethemachine.ai/checkout.md. The full site index (every tool, model and page, about 82 KB): https://insidethemachine.ai/llms.txt · JSON: https://insidethemachine.ai/index.json