Quick answer
An open-weight model is an AI model whose trained weights are available to download or run outside the vendor's hosted product. It may still have a restrictive custom license.
The practical question is not only "can I download it?" but also "what does the license allow me to do with it?"
The three labels we use
- Open source
- Weights are available under a recognized open-source-style license such as Apache-2.0 or MIT. This is the clearest category for reuse, though you should still read the exact license.
- Open weights
- Weights are downloadable, but the license may include custom restrictions, commercial conditions, redistribution limits, or acceptable-use terms.
- Proprietary
- The model is primarily accessed through a hosted product, API, or private deployment path, and weights are not generally released.
Current catalog mix: 95 open-source, 96 open-weight, and 113 proprietary models.
Commercial-use checklist
Before you build on a downloadable model, check these points in the primary source or license:
- Whether commercial use is explicitly allowed.
- Whether there are user-count, revenue, geography, field-of-use, or competitor restrictions.
- Whether redistribution, fine-tuning, distillation, or hosting the model for others is allowed.
- Whether attribution, model naming, or disclosure is required.
- Whether a separate paid license is needed above a certain scale.
This guide is not legal advice; it is a reading checklist for model release pages and licenses.
Recent downloadable examples
| Model | Lab | Catalog label | License | Params | Released |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash-0731 | DeepSeek | Open source | MIT | 284B · 13B active | Jul 31, 2026 |
| Ling-3.0-flash | inclusionAI | Open weights | Apache 2.0 (announced; weights not yet posted as of 2026-07-24) | 124B · 5.1B active | Jul 23, 2026 |
| Laguna S 2.1 | Poolside | Open source | OpenMDW-1.1 | 118B · 8B active | Jul 21, 2026 |
| Kimi K3 | Moonshot | Open weights | Modified MIT | 2.8T · 50B active | Jul 16, 2026 |
| Inkling | Thinking Machines | Open source | Apache-2.0 | 975B · 41B active | Jul 15, 2026 |
| Hunyuan Hy3 | Hunyuan | Open source | Apache-2.0 | 295B · 21B active | Jul 6, 2026 |
| Nemotron-Labs-3-Puzzle-75B-A9B | NVIDIA | Open weights | OpenMDW-1.1 | 75.3B · 9.3B active | Jul 6, 2026 |
| Mistral frontier open-weight MoE (unnamed) | Mistral | Open weights | Undisclosed | Undisclosed | Jul 6, 2026 |
Frequently asked questions
Is open weight the same as open source?
No. Open weight means the trained weights are available. Open source usually implies a recognized permissive or copyleft license, and may also include code, data, or training details.
Can I use an open-weight model commercially?
Sometimes, but not always. Commercial use depends on the model license, acceptable-use terms, scale restrictions, attribution requirements, and any separate service terms. For real deployment decisions, read the license text linked from the model page.
Why does LLM Releases separate open source and open weights?
Because downloadable weights are not enough to know how a model can be reused. A custom model license can allow local use while restricting redistribution, commercial use, competitors, or high-scale deployments.
Where to go next
Browse the live open-weight model catalog, compare open coding models, or check the methodology for how sources are reviewed.