Verified model directory

Open-Weight Models Directory

32 model passports. One consistent evidence model.

Compare architecture, context, licensing, hardware class and deployment routes without flattening provider-specific facts into the base model.

Model factsArchitecture, modalities, parameters and context.
License factsCommercial use, modification, redistribution and special terms.
Deployment factsRuntimes, provider envelopes and hardware evidence.
EvidencePrimary sources and dated verification.
All models

Browse the registry.

Use the search box to find a model, developer, license family or hardware class. Every row links to a full Model Passport.

32 models
ModelParametersContextLicenseHardware classPassport
DeepSeek-R1DeepSeek671B128K tokensMIT>80B · datacenter / multi-GPUPassport →
DeepSeek-V3.1DeepSeek~671B-class128K tokensMIT>80B · datacenter / multi-GPUPassport →
Devstral Small 2505Mistral AI24B128K tokensApache 2.017–32B · high-memory workstationPassport →
Gemma 3 12B ITGoogle DeepMind12B128K tokensGemma Terms9–16B · workstation/localPassport →
Gemma 3 27B ITGoogle DeepMind27B128K tokensGemma Terms17–32B · high-memory workstationPassport →
Gemma 3 4B ITGoogle DeepMind4B128K tokensGemma Terms≤8B · consumer/localPassport →
Gemma 3n E4B ITGoogle DeepMind~8B raw / E4B effective profile32K tokensGemma Terms≤8B · consumer/localPassport →
GLM-4.5Z.ai355B128K tokensMIT>80B · datacenter / multi-GPUPassport →
GLM-4.5 AirZ.ai106B128K tokensMIT>80B · datacenter / multi-GPUPassport →
gpt-oss-120bOpenAI116.8B131,072 tokensApache 2.0>80B · datacenter / multi-GPUPassport →
gpt-oss-20bOpenAI20.9B131,072 tokensApache 2.017–32B · high-memory workstationPassport →
Granite 3.3 8B InstructIBM8B128K tokensApache 2.0≤8B · consumer/localPassport →
Kimi K2 InstructMoonshot AI1T128K tokensModified MIT>80B · datacenter / multi-GPUPassport →
Llama 3.3 70B InstructMeta70B128K tokensLlama 3.3 Community33–80B · large-memory / multi-GPUPassport →
Llama 4 MaverickMeta~400B1000000 tokensLlama 4 Community>80B · datacenter / multi-GPUPassport →
Llama 4 ScoutMeta~109B10,000,000 tokensLlama 4 Community>80B · datacenter / multi-GPUPassport →
Magistral Small 2506Mistral AI24B128K nominal · 40K recommended for qualityApache 2.017–32B · high-memory workstationPassport →
Mathstral 7B v0.1Mistral AI7B32,768 tokensApache 2.0≤8B · consumer/localPassport →
Mistral Nemo Instruct 2407Mistral AI / NVIDIA12B128K tokensApache 2.09–16B · workstation/localPassport →
Mistral Small 3.1 24B InstructMistral AI24B128K tokensApache 2.017–32B · high-memory workstationPassport →
OLMo 2 13B InstructAi213B4,096 tokensApache 2.09–16B · workstation/localPassport →
OLMo 2 32B InstructAi232B4,096 tokensApache 2.017–32B · high-memory workstationPassport →
Phi-4Microsoft14B16K tokensMIT9–16B · workstation/localPassport →
Phi-4 Mini InstructMicrosoft3.8B128K tokensMIT≤8B · consumer/localPassport →
Phi-4 Multimodal InstructMicrosoft5.6B128K tokensMIT≤8B · consumer/localPassport →
Qwen2.5-Coder-32B-InstructQwen / Alibaba~32.5B32K native · long-context extension documented by QwenApache 2.017–32B · high-memory workstationPassport →
Qwen2.5-VL-72B-InstructQwen / Alibaba~72B32K recommended default · config exposes 128K positionsQwen License33–80B · large-memory / multi-GPUPassport →
Qwen3-235B-A22BQwen / Alibaba235B32,768 native · 131,072 with YaRNApache 2.0>80B · datacenter / multi-GPUPassport →
Qwen3-30B-A3BQwen / Alibaba30.5B32,768 native · 131,072 with YaRNApache 2.017–32B · high-memory workstationPassport →
Qwen3-32BQwen / Alibaba32.8B32,768 native · 131,072 with YaRNApache 2.017–32B · high-memory workstationPassport →
Qwen3-Coder-480B-A35B-InstructQwen / Alibaba480B256K tokensApache 2.0>80B · datacenter / multi-GPUPassport →
SmolLM3 3BHugging Face3B64K native · up to 128K extendedApache 2.0≤8B · consumer/localPassport →
Explore by intent

Move from model names to decisions.

These paths regroup the same verified records around deployment and licensing questions people actually search for.

FAQ

How to read the directory.

What is an open-weight model directory?

It is a structured index of models whose trained weights are obtainable, with separate records for architecture, licensing and deployment evidence.

Are all models in this directory open source?

No. Open weight describes weight availability. Some entries use Apache 2.0 or MIT, while others use custom model terms such as the Llama Community License, Gemma Terms or the Qwen License.

Can every model be self-hosted?

Weight access can make independent deployment possible, but practical feasibility varies from consumer-class 3B models to datacenter-scale mixture-of-experts models.

How are model facts verified?

OpenWeightModels prioritizes exact model cards, license files, official repositories and named deployment-provider documentation, with a visible verification date.