stinkydonkey

stinky
donkey

open weights ยท runs local

AI that lives on your machine. A field guide to the open models worth running, the tools to run them, and how privacy people actually put them to work.

Directory

Open models worth knowing

Model families you can download and run yourself. "Open" means very different things depending on the license, so every entry is tagged.

  • Fully open weights, data and training code
  • Open license Apache 2.0 or MIT weights
  • Open weights custom license, read the terms

OLMo

Ai2

The purist's pick. Weights, training data, code and logs are all public, so you can check exactly what went in.

Fully openLaptop to workstation
huggingface.co/allenai

Gemma 4

Google

Small edge sizes up to a 31B model. Reads text, images and audio. A strong default for laptops and phones.

Apache 2.0Phone to workstation
huggingface.co/google

gpt-oss

OpenAI

20B and 120B reasoning models. The 20B fits in about 16GB of memory, which makes it a solid local all-rounder.

Apache 2.016GB laptop and up
huggingface.co/openai

Qwen

Alibaba

The widest size range in open AI, from tiny models to giants. The 27B is a local favorite for coding and agents.

Apache 2.0, most sizesLaptop to server
huggingface.co/Qwen

DeepSeek V4

DeepSeek

Frontier-class mixture-of-experts models under a permissive license. Built for serious hardware or a self-hosted server.

MITMulti-GPU server
huggingface.co/deepseek-ai

Mistral

Mistral AI

Mistral Small is fast and capable on a single good GPU. The bigger Medium model ships under its own terms.

Apache 2.0, SmallWorkstation
huggingface.co/mistralai

Phi

Microsoft

Small models that punch above their size. Good for older laptops, edge boxes and quick offline tasks.

MITLaptop
huggingface.co/microsoft

GLM

Z.ai

Strong at coding and agent work. The Flash models are MIT licensed. The flagship carries its own license.

MIT, FlashWorkstation to server
huggingface.co/zai-org

Kimi

Moonshot AI

Very large agentic models with long context. Free to run, but big commercial deployments owe attribution and an agreement.

Custom licenseServer cluster
huggingface.co/moonshotai

Llama 4

Meta

The model that kicked off the open wave. Still widely supported, though its community license adds attribution and usage caps.

Community licenseServer
huggingface.co/meta-llama

Licenses and versions move fast. Always read the license on the model card before you build on it. Directory last checked October 2026.

Directory

AI by category: own it or rent it

For every job, the open tools you can run yourself, next to the big cloud services so you know what you're comparing against.

About the cloud column: these services run on someone else's servers, so your prompts, files and outputs pass through them. Terms, prices and even the product itself can change overnight. OpenAI shut down its Sora video app in 2026. Listed for comparison, not endorsement.

Writing and chat

Drafting, editing, brainstorming and everyday questions.

Run it yourself

  • Any model in the list above

    Gemma, gpt-oss, Qwen and friends all write well. Start with whatever fits your hardware.

    See the models
  • Open WebUIopen source

    A ChatGPT-style interface for your own models, on your own server.

  • Janopen source

    Offline desktop chat app. Download a model and go.

Cloud versions

  • ChatGPTOpenAI

    The best-known chat assistant.

    Cloud
  • ClaudeAnthropic

    Chat assistant known for long documents and writing.

    Cloud
  • GeminiGoogle

    Google's assistant, tied into its apps.

    Cloud
  • GrokxAI

    Assistant built into X.

    Cloud

Coding

Autocomplete, code review and agents that make changes. Pair any of these with a local model.

Run it yourself

  • Continueopen source

    Copilot-style assistant for VS Code and JetBrains, pointed at your own model.

    Apache 2.0
  • Aideropen source

    Pair programming in the terminal. Edits files and commits to git.

    Apache 2.0
  • Clineopen source

    Coding agent inside VS Code that plans and runs multi-step changes.

    Apache 2.0
  • OpenHandsopen source

    Autonomous software agent you can self-host.

    MIT

Cloud versions

  • GitHub CopilotGitHub / Microsoft

    The most widely used coding assistant.

    Cloud
  • CursorAnysphere

    AI-first code editor.

    Cloud
  • Claude CodeAnthropic

    Coding agent for the terminal and IDE.

    Cloud
  • CodexOpenAI

    OpenAI's coding agent.

    Cloud

Images

Text to image, editing and design work.

Run it yourself

  • Qwen-ImageAlibaba

    Strong all-rounder, good with text inside images. Edit variants too.

    Apache 2.0
  • FLUX.2Black Forest Labs

    Top-tier quality. The small klein versions fit around 13GB of VRAM. The full dev model needs a paid license for business use.

    Check terms
  • Stable Diffusion 3.5Stability AI

    Huge ecosystem of fine-tunes and add-ons.

    Community license
  • ComfyUIopen source app

    Node-based workspace for running image and video models locally.

    GPL-3.0
  • InvokeAIopen source app

    Friendlier canvas-style app for local image work.

    Apache 2.0

Cloud versions

  • MidjourneyMidjourney

    Known for its distinctive artistic look.

    Cloud
  • ChatGPT imagesOpenAI

    Image generation and editing inside ChatGPT.

    Cloud
  • Gemini imagesGoogle

    Image generation and editing inside Gemini.

    Cloud
  • FireflyAdobe

    Built into Photoshop and Adobe's apps.

    Cloud

Video

Text and image to video. The hungriest category for GPU memory.

Run it yourself

  • Wan 2.2Alibaba

    The most open of the strong video models. The 5B version wants about 24GB of VRAM.

    Apache 2.0
  • LTX-2Lightricks

    Fast and runs on 16GB. Free for companies under about $10M in revenue.

    Community license
  • HunyuanVideo 1.5Tencent

    High quality at around 14GB with offloading. License excludes the EU, UK and South Korea.

    Community license
  • Mochi 1Genmo

    Smooth, realistic motion at about 20GB.

    Apache 2.0

Cloud versions

  • VeoGoogle

    Google's flagship video model.

    Cloud
  • RunwayRunway

    Video generation and editing suite for creators.

    Cloud
  • KlingKuaishou

    Popular for realistic clips.

    Cloud
  • Dream MachineLuma AI

    Text and image to video.

    Cloud

Voice

Text to speech and voice cloning. Only clone voices you have permission to use.

Run it yourself

  • Kokorohexgrad

    Tiny model that runs on a plain CPU with dozens of preset voices.

    Apache 2.0
  • ChatterboxResemble AI

    Clones a voice from a few seconds of audio, 23 languages, built-in watermarking.

    MIT
  • Qwen3-TTSAlibaba

    Small sizes, voice design and cloning, 10 languages.

    Apache 2.0
  • ZonosZyphra

    High-fidelity cloning that runs on about 6GB of VRAM.

    Apache 2.0

Cloud versions

  • ElevenLabsElevenLabs

    The best-known voice and cloning platform.

    Cloud
  • OpenAI voicesOpenAI

    Speech built into ChatGPT and the API.

    Cloud

Transcription

Speech to text for calls, interviews and voice notes.

Run it yourself

  • WhisperOpenAI

    The standard open speech recognition model, dozens of languages.

    MIT
  • whisper.cppopen source

    Runs Whisper fast on laptops, even without a GPU.

    MIT
  • ParakeetNVIDIA

    Very fast, accurate English transcription.

    CC BY 4.0

Cloud versions

  • OtterOtter.ai

    Meeting notes and transcripts.

    Cloud
  • DescriptDescript

    Transcribe and edit audio and video by editing text.

    Cloud
  • RevRev

    AI and human transcription service.

    Cloud

Music

Songs, beats and sound from a text prompt.

Run it yourself

  • ACE-StepACE Studio and StepFun

    Fast full-song generation with vocals.

    Apache 2.0
  • YuEM-A-P

    Lyrics to full songs with vocals and backing.

    Apache 2.0
  • MusicGenMeta

    Instrumental clips from text or a melody.

    Non-commercial

Cloud versions

  • SunoSuno

    The most popular AI song generator.

    Cloud
  • UdioUdio

    Song generation, now working under record label licensing deals.

    Cloud

3D

3D models from text or a single image.

Run it yourself

  • TRELLISMicrosoft

    Image or text to textured 3D assets.

    MIT
  • Hunyuan3DTencent

    High-quality shapes and textures from an image.

    Community license

Cloud versions

  • MeshyMeshy

    Text and image to 3D for games and printing.

    Cloud

Research and your docs

Search the web or chat with your own files and get cited answers.

Run it yourself

  • AnythingLLMopen source

    Desktop or server app for chatting with your documents privately.

    MIT
  • PrivateGPTopen source

    Ask questions about your files with nothing leaving your machine.

    Apache 2.0
  • Perplexicaopen source

    Self-hosted answer engine for web search with sources.

    MIT

Cloud versions

  • PerplexityPerplexity AI

    AI search with citations.

    Cloud
  • NotebookLMGoogle

    Chat with uploaded documents and generate audio summaries.

    Cloud

Licenses summarized for quick comparison, not legal advice. Read the actual license before commercial use. Last checked October 2026.

Run it yourself

Tools to run models at home

No accounts, no API keys, no cloud. Pick one, download a model, and your prompts never leave the room.

OllamaOne command to pull and run a model. The easiest start. LM StudioA desktop app with a model browser and chat window. JanAn open source, offline-first desktop chat app. llama.cppThe engine under much of local AI. Runs on almost anything. Open WebUIA self-hosted chat interface for your home server. vLLMHigh-throughput serving when you outgrow a single laptop.

Field notes

How cypherpunks are probably using AI

Nobody publishes a list, which is kind of the point. These are the patterns that make sense when privacy comes first.

Air-gapped assistants

A capable model on a machine that never touches the internet. Ask it anything and nothing gets logged anywhere else.

Chat with your own files

Point a local model at notes, PDFs and encrypted archives. Search and summarize without uploading a single page.

Auditing code before trusting it

A second set of eyes on wallets, smart contracts and open source dependencies before anything gets signed or installed.

Reading the fine print

Feeding privacy policies and terms of service to a model to flag what an app really collects and who it shares with.

Private transcription and translation

Turning calls, voice memos and foreign-language docs into text locally, with no audio sent to a cloud service.

Self-hosted coding copilots

An autocomplete and code helper that runs on your own GPU, so proprietary code never leaves the building.

Verifying the model itself

Checking hashes, pinning versions and preferring fully open models, because a model you can't inspect is one more thing to trust.

Sharing compute, not data

Pooling GPUs with friends over private networks and running inference on hardware the group owns.

Next up

Our own open model

We're building a Stinky Donkey model of our own. Open source from day one, made to run on hardware you actually own.