Weights available
You can obtain the trained parameters and run the model yourself, subject to its license and technical requirements.
- Self-managed inference possible
- License varies
- Training data may remain unavailable
Independent · source-first · machine-readable
Open Weight Models
A practical reference for AI models whose trained weights are available to download and run.
Understand what “open weight” actually means, compare model families and licenses, estimate deployment requirements, and follow primary sources instead of relying on ambiguous labels.
25 selected familiesPrimary sources attachedLast verified 26 Sep 2026
Open weights ≠ automatically open sourceWeight access describes one layer of openness; code, data information and legal rights vary.
Licenses are model-specificApache 2.0, MIT, community licenses and modified terms can imply very different rights.
Deployment is a systems questionParameter count alone does not determine memory, throughput or cost.
Sources beat summariesEvery registry entry points back to a publisher-controlled model card or repository.
Definition
What is an open-weight model?
An open-weight model is an AI model whose trained parameters — the learned numerical values commonly called weights — are made available so that others can obtain and run the model outside the publisher’s hosted API.
The term focuses on access to the trained parameters. It does not, by itself, say that the training dataset is public, that the complete training code is available, or that every use is permitted. Those questions must be answered separately from the model card, license and accompanying policies.
What are “weights” technically?
During training, a neural network adjusts very large collections of numeric parameters. Those learned values are stored in tensors and checkpoint files. At inference time, the architecture plus these parameters transform inputs into outputs. Making the weights available lets another operator load the trained model on compatible infrastructure without calling the original provider’s API.
Why “open weight” and “open source” are not synonyms
The Open Source Initiative’s Open Source AI Definition is broader: it requires the freedoms to use, study, modify and share, and describes a preferred form for modification that includes data information, code and parameters. A model can therefore have downloadable weights without necessarily satisfying the OSI definition of Open Source AI.
Broader openness
Under the OSI definition, meaningful freedoms and the preferred form for modification extend beyond final weights.
- Use, study, modify, share
- Code and data information matter
- OSI-approved terms are relevant
Hosted access
The trained parameters are not distributed to users. Access is typically provided through an API or hosted product.
- Provider controls deployment
- No direct weight access
- Operational control is narrower
TrainingData + objectiveExamples, mixtures, preprocessing, optimization
ArchitectureNetwork designLayers, attention, experts, tokenizer
WeightsLearned parametersThe trained numeric state distributed as checkpoints
RuntimeInference engineLoads and executes the model
ApplicationYour systemAgents, assistants, RAG, code, multimodal workflows
Why it matters
Open weights move deployment choices back to the operator.
They can enable control, portability and customization — but they also move infrastructure, security, evaluation and compliance responsibilities onto the deployer.
Infrastructure control
Run on workstations, private clouds, data centers or edge hardware when the model and runtime support it.
Customization
Fine-tune, adapt, quantize or package a model using compatible tools and within the rights granted by its terms.
Data locality
Self-hosting can keep prompts, retrieved documents and outputs inside infrastructure chosen by the operator.
Runtime choice
Choose serving software, quantization formats, accelerators and scaling strategies instead of a single API stack.
Model registry
Selected open-weight model families
This is a curated 25-family index, not a leaderboard. Model-specific checkpoints can differ in license, context, modality and hardware requirements.
25 families
| Family | Developer | Size | Context | License | Focus | Sources |
|---|---|---|---|---|---|---|
| gpt-ossgpt-oss-20b · gpt-oss-120b | OpenAI | 20B / 120B class | See official model card | Apache 2.0 + usage policy | Reasoning, tool use, self-hosted inference | Primary ↗ Hub ↗ |
| Llama 4Scout 17B-16E · Maverick 17B-128E | Meta | 17B active class | Scout 10M · Maverick 1M | Llama 4 Community License | General-purpose, multimodal, long context | Primary ↗ Hub ↗ |
| Qwen30.6B–32B dense · 30B-A3B · 235B-A22B | Qwen / Alibaba | 0.6B–235B total | Up to 256K; selected variants extend further | Apache 2.0 | Reasoning, coding, multilingual, agents | Primary ↗ Hub ↗ |
| Mistral Small 4Mistral-Small-4-119B-2603 | Mistral AI | 119B total · 6.5B active/token | 256K | Apache 2.0 | Reasoning, vision, agents, multilingual | Primary ↗ Hub ↗ |
| DeepSeek-V3.2DeepSeek-V3.2 | DeepSeek | 685B parameters | See official model card | MIT | Reasoning, coding, tool use, agentic AI | Primary ↗ Hub ↗ |
| Kimi K2Kimi-K2-Base · Instruct · Thinking | Moonshot AI | 1T total · 32B active | 128K (K2 base card) | Modified MIT | Agentic intelligence, coding, reasoning | Primary ↗ Hub ↗ |
| GLM-4.5GLM-4.5 · GLM-4.5-Air | Z.ai | 355B-A32B · 106B-A12B | See official model card | MIT | Agents, reasoning, coding, tool use | Primary ↗ Hub ↗ |
| Phi-4Phi-4 · Phi-4 mini · reasoning · multimodal | Microsoft | 14B base Phi-4; family includes smaller variants | 16K on Phi-4 model card | MIT | Reasoning, constrained compute, multimodal variants | Primary ↗ Hub ↗ |
| Granite 4.2Granite 4.2 3B · 8B · 30B | IBM | 3B · 8B · 30B | 128K native; longer extension documented | Apache 2.0 | Enterprise, reasoning, code, tools, multilingual | Primary ↗ Hub ↗ |
| Gemma 3Gemma 3 family | Google DeepMind | Multiple sizes | See official model card | Gemma Terms | Multimodal, multilingual, local / on-device use | Primary ↗ Hub ↗ |
| MiniMax-M2MiniMax-M2 | MiniMax | See official model card | See official model card | Modified MIT | General-purpose, coding, agentic workloads | Primary ↗ Hub ↗ |
| Nemotron 3 NanoNVIDIA-Nemotron-3-Nano-4B | NVIDIA | 4B class | See official model card | NVIDIA Nemotron Open Model License | Local inference, compact deployment, NVIDIA ecosystem | Primary ↗ Hub ↗ |
| DeepSeek-R1DeepSeek-R1 · DeepSeek-R1-Zero | DeepSeek | 671B total · 37B active | 128K | MIT | reasoning · math · coding · distillation | Primary ↗ Hub ↗ |
| Qwen3-CoderQwen3-Coder-480B-A35B-Instruct | Qwen / Alibaba | 480B total · 35B active | 256K native · extendable to 1M | Apache 2.0 | agentic coding · repository understanding · tool use | Primary ↗ Hub ↗ |
| Qwen3-VLQwen3-VL-235B-A22B-Instruct · Qwen3-VL Thinking variants | Qwen / Alibaba | Up to 235B total · 22B active | 256K native · expandable to 1M | Apache 2.0 | vision-language · visual agents · video · OCR · spatial reasoning | Primary ↗ Hub ↗ |
| OLMo 3OLMo 3 7B · OLMo 3 32B · Instruct · Think | Ai2 | 7B · 32B | 65,536 | Apache 2.0 | fully open research · reasoning · instruction following | Primary ↗ Hub ↗ |
| SmolLM3SmolLM3-3B · SmolLM3-3B-Base | Hugging Face | 3B | 64K trained · up to 128K with YaRN | Apache 2.0 | small models · reasoning · multilingual · local inference · tool use | Primary ↗ Hub ↗ |
| Command Ac4ai-command-a-03-2025 | Cohere Labs | 111B | 256K supported | CC BY-NC 4.0 + AUP | enterprise · agents · multilingual · RAG · tool use | Primary ↗ Hub ↗ |
| Falcon-H1Falcon-H1-34B-Instruct · Falcon-H1-34B-Base | Technology Innovation Institute | 34B | See official model card | Falcon LLM License | multilingual · general-purpose · efficient sequence modeling | Primary ↗ Hub ↗ |
| InternLM3InternLM3-8B-Instruct | InternLM | 8B | See official model card | Apache 2.0 | general-purpose · reasoning · knowledge-intensive tasks | Primary ↗ Hub ↗ |
| Yi-1.5Yi-1.5-6B · Yi-1.5-9B · Yi-1.5-34B | 01.AI | 6B · 9B · 34B | 4K · 16K · 32K variants | Apache 2.0 | general-purpose · coding · math · reasoning | Primary ↗ Hub ↗ |
| Hunyuan-A13BHunyuan-A13B-Instruct | Tencent | A13B family | See official model card | Tencent Hunyuan Community License⚠ license caveat | general-purpose · reasoning · tool use | Primary ↗ Hub ↗ |
| ERNIE 4.5ERNIE-4.5-300B-A47B-Base | Baidu | 300B total · 47B active | 131,072 | Apache 2.0 | general language · long context · multimodal family variants | Primary ↗ Hub ↗ |
| DevstralDevstral Small 1.1 | Mistral AI / All Hands AI | 24B | 128K | Apache 2.0 | agentic coding · software engineering · repository editing | Primary ↗ Hub ↗ |
| MagistralMagistral Small 1.0 | Mistral AI | 24B | 128K advertised · 40K recommended for best performance | Apache 2.0 | reasoning · math · multilingual · local deployment | Primary ↗ Hub ↗ |
Registry records are informational. Always verify the exact checkpoint’s model card, license and deployment documentation before commercial or production use.
Licensing
“Open” is not a license.
A project should record the exact license and policy attached to the checkpoint, not infer rights from marketing terminology.
Apache 2.0 / MIT
These can grant broad rights, including commercial use, modification and redistribution, but the full model release may still include separate usage policies or third-party components.
Model-specific licenses
Llama, Gemma and other families use terms tailored to their ecosystems. Read the actual agreement rather than assuming equivalence with standard OSS licenses.
Near-standard ≠ identical
“Modified MIT” and similar labels should be treated as distinct legal documents. The differences may be operationally important.
Deployment rule: Record the exact model identifier, exact license URL and verification date together. A family name alone is not enough for compliance.
Hardware & quantization
Can you actually run the model?
That question depends on more than parameter count. Precision, active parameters, KV cache, context length, batch size, runtime buffers and offloading all matter.
First-order estimateraw weight memory ≈ parameters × bytes per parameter
FP16/BF16 ≈ 2 BINT8 ≈ 1 B4-bit ≈ 0.5 B*
VRAM is not just weights
Inference also needs memory for KV cache, temporary buffers, runtime overhead and sometimes multiple replicas.
MoE changes the picture
Mixture-of-Experts models can activate only part of the network per token, but the full checkpoint may still need to be stored or distributed.
Quantization is a trade-off
Lower precision can reduce memory and improve accessibility, while quality and speed depend on method, hardware and runtime support.
* Approximation for raw weight storage only; real deployment memory is higher.
Inference
The runtime is part of the model decision.
Open-weight models can often be served through multiple stacks. Compatibility varies by architecture and checkpoint.
vLLM
High-throughput serving for many transformer architectures with OpenAI-compatible APIs and accelerator-focused execution.
SGLang
A serving and programming stack used by multiple current open-weight model releases for reasoning and tool workflows.
llama.cpp
Portable local inference with broad support for GGUF quantizations and CPU/GPU execution.
Ollama
A developer-friendly local model workflow that packages and serves many popular open-weight model families.
🤗 Project ecosystem
Open Weight Models on Hugging Face
GitHub Pages is the long-form reference and machine-readable index. Hugging Face is the companion workspace for interactive Spaces, curated Collections, datasets and model-infrastructure experiments.
For search, agents & tools
Human-readable and machine-readable by design.
Good GEO is not keyword stuffing. It means clear definitions, structured entities, primary-source links, stable URLs and data that systems can parse reliably.
/data/models.jsonStructured model registryFamily-level metadata with licenses, focus, sources and verification dates.
/llms.txtMachine-oriented site guideA concise map of canonical resources and editorial principles.
/sitemap.xmlIndexable resource mapCanonical pages and machine-readable assets for discovery.
FAQ
Questions people ask about open-weight models
Are open-weight models the same as open-source models?
No. “Open weight” means the trained parameters are available. Open Source AI is a broader claim about freedoms and the preferred form for modification. The OSI definition includes requirements involving data information, code and parameters.
Can I use an open-weight model commercially?
Sometimes, but not because it is open-weight. Commercial rights depend on the exact license and any additional policies attached to the checkpoint.
Can I run open-weight models locally?
Many can run locally, particularly smaller or quantized checkpoints. Feasibility depends on model size, precision, architecture, context length, runtime and your available RAM/VRAM.
Do open weights make a model private?
No. They give you the option to self-host. Privacy then depends on how you deploy, log, network and secure the system.
What is the difference between a model family and a checkpoint?
A family groups related releases. A checkpoint is a particular set of weights with its own identifier, size, precision, training state and sometimes distinct license or usage terms.
Why does this project link to Hugging Face?
Hugging Face is a major distribution and tooling hub for open-weight models. The companion open-weight organization can host interactive explorers, collections and structured datasets while this site remains the canonical long-form reference.
Is this a benchmark leaderboard?
No. The project prioritizes operational questions: access, licensing, formats, hardware, inference and deployment. Benchmarks can be useful evidence, but no single score represents overall model quality.
Primary references
Trace claims back to publishers.
The registry is designed to summarize, not replace, model cards and license documents.
OpenAIgpt-oss open models ↗
MetaLlama models repository ↗
QwenQwen3 repository ↗
Mistral AIMistral Small 4 ↗
DeepSeekDeepSeek-V3.2 ↗
Moonshot AIKimi K2 ↗
Z.aiGLM-4.5 ↗
OSIOpen Source AI Definition ↗
Cite this reference
A stable citation block makes the project easier to reference in articles, repositories and research notes.
@misc{openweightmodels2026,
title = {Open Weight Models},
author = {{Open Weight Models}},
year = {2026},
howpublished = {\url{https://openweightmodels.eu/}},
note = {Independent reference for open-weight AI models; accessed 2026-09-26}
}