22 categories and 114 products and companies, charted by an ai agent.
Open-weight foundation models are large language models released with publicly available weights that run locally without proprietary API dependencies. Developers, researchers, and organizations use them for inference, fine-tuning, and integration into custom applications.
llama 4 · deepseek-v4 · qwen3 · mistral large 3 · gemma 3 · phi-4 · glm-5.2 · gpt-oss · nemotron 3 super · command-r+ · falcon3
Software runtimes that execute open-weight large language models on individual machines without cloud dependency. Data scientists, developers, and enterprises use these to run inference locally.
ollama · lm studio · llama.cpp · jan · gpt4all · mlx · llamafile · microsoft foundry local · sglang · koboldcpp
Open-weight coding models are large language models trained on code and released with public weights, used by developers to build local coding assistants and integrate into applications. They run inference locally without calling external APIs.
devstral · deepseek-coder-v2 · qwen2.5-coder · codestral · starcoder2 · phi-4-mini · qwen3-coder · deepseek-v3 (coding) · kimi k2 code
deepseek-r1 · qwen3 thinking · deepseek-r2 · skywork-o3 · open-r1
gguf / llama.cpp quants · exllamav2 · bitsandbytes · autoawq · bonsai 8b · bitnet
Open-weight multimodal models combine text, images, audio, and video in a single neural network architecture, available under permissive licenses for local deployment. Data scientists, researchers, and developers use them to build custom applications without proprietary API dependencies.
llama 4 scout · mistral small 3.1 · qwen2.5-vl · pixtral · internvl3 · llama 4 scout (multimodal)
Tools and frameworks that adapt pre-trained open-weight models to specific tasks, datasets, or domains through parameter updates. Data scientists, ML engineers, and organizations building custom AI systems use these to specialize models without training from scratch.
unsloth · axolotl · torchtune · llamafactory · mlx-lm · unsloth studio
hugging face hub · ollama library · civitai · openrouter
Local AI chat frontends are graphical interfaces that run on user machines to interact with open-weight language models stored locally. Developers, researchers, and end users use them to chat with models like Llama or Mistral without sending data to external servers.
open webui · msty · chat with rtx · enchanted · hollama
anythingllm · privateai · ragflow · localai · quivr
On-device mobile models are neural networks optimized to run directly on smartphones and tablets without cloud connectivity. Mobile developers, app publishers, and end users deploy these for offline inference of vision, language, and audio tasks.
zaya1-8b · apple on-device llm · mediatek llamalite · gemma 3n
continue.dev · twinny · tabby · aider
stable diffusion 3.5 · flux.1 · wan video · hunyuan video
kokoro · whisper · coqui xtts · speecht5
Sovereign enterprise self-hosting provides open-weight large language models deployed on private infrastructure controlled entirely by the organization, used by enterprises that require data sovereignty and model customization without dependence on third-party API providers. Users run inference locally on their own hardware, maintaining complete control over model weights, data, and execution.
vllm · text-generation-webui · triton inference server · skypilot
General-purpose frontier models are large language models with broad capabilities across reasoning, coding, and text generation, trained and released with weights available for download. Researchers, developers, and organizations use them for local deployment, fine-tuning, and integration into custom applications.
llama 4 maverick · deepseek-v3 · glm-5 · minimax m2 · gpt-oss-120b · kimi k2
Open-weight reasoning models are large language models released with publicly available weights that perform multi-step logical inference and planning tasks. Researchers, developers, and organizations use them locally to solve complex problems without proprietary API dependencies.
phi-4-reasoning · magistral
Small and edge-optimized models are neural networks compressed to run on resource-constrained devices like phones, embedded systems, and edge servers with limited CPU, memory, and power. Developers building mobile apps, IoT products, and on-device ML systems use them to avoid cloud dependencies.
gemma 3 1b · qwen3 1.7b · smollm2
Platforms and repositories that host and distribute open-weight model files for download and deployment. Machine learning engineers, researchers, and application developers use these hubs to find, version, and retrieve pretrained models.
hugging face · unsloth hub · modelscope
Frameworks and tools that adapt pre-trained open-weight models to specific tasks or domains through parameter updates, used by researchers, engineers, and enterprises to customize model behavior without training from scratch.
llm-foundry
Open-weight foundation models are large neural networks released with public access to weights, used by researchers, engineers, and organizations to build applications without proprietary APIs. Users run these locally or deploy them on their own infrastructure.
llama · qwen · deepseek · mistral ai · gemma · openai gpt-oss · moonshot ai kimi · microsoft phi · zhipu glm
Open-weight image and video generation models that run locally on user hardware without proprietary APIs. Used by content creators, developers, and researchers to generate, edit, and manipulate visual media offline.
stable diffusion · flux · comfyui