Open-Weight AI Models and Their Licences, November 2026
18 notable open-weight models were released in our tracker period. 8 use MIT or Apache 2.0, the most permissive licences; 3 carry limits such as non-commercial use, revenue caps or excluded regions. Newest first, with each model's size, context window and licence.
Updated 13 Nov 2026 from the AI model release tracker.
| Released | Model | Developer | Context | Licence |
|---|---|---|---|---|
| 5 Nov 2026 | ERNIE 5.5 Flash Base (280B total, 14B active) | Baidu | 128K training, extendable to 1M tokens | ERNIE 5.5 Flash Base license; excludes EU, UK, South Korea |
| 4 Nov 2026 | Kimi K3.1 Pro (2.8T total, about 104B active) | Moonshot AI | About 1M tokens | Kimi K3.1-Pro License |
| 29 Oct 2026 | Granite Switch 4.1.1 (8B and 30B) | IBM | 131,072 tokens | Apache 2.0 |
| 29 Oct 2026 | Mistral Large 4 (about 1T total, 52B active) | Mistral AI | 1M tokens | Mistral Ecosystem Open-Weight license, separate license for Model-as-a-Service |
| 28 Oct 2026 | Nemotron 3.5 Transcribe (30B total, about 4B active) | NVIDIA | not stated | OpenMDW-1.1 |
| 27 Oct 2026 | LFM3 (six models of about 2B parameters) | Liquid AI | 256K tokens (text model) | LFM Open License v1.0, commercial use under $10M revenue |
| 27 Oct 2026 | Ling-3.1-flash-1019 (560B total, 25B active) | Ant Group (inclusionAI) | 1M tokens | MIT |
| 27 Oct 2026 | Unlimited OCR 2 (about 3.3B) | Baidu | not stated | MIT |
| 17 Oct 2026 | DeepSeek-V4.1-Flash-1017 | DeepSeek | 1M tokens | MIT |
| 16 Oct 2026 | Beam 1.0 (501B total, 23B active) | Reflection AI | 1M tokens | Apache 2.0 |
| 15 Oct 2026 | Nemotron 3.5 Super VL (120B total, 12B active) | NVIDIA | Up to 1M tokens | OpenMDW 1.1 |
| 15 Oct 2026 | Nemotron 3.5 Bolt 74B-A9B 24GB | NVIDIA | 256K tokens | OpenMDW 1.1 |
| 15 Oct 2026 | Muse Spark Open (816B total, 42B active) | Meta | 1M tokens | Muse Model License |
| 3 Oct 2026 | Kolibri-1 (78B total, 3.46B active) | Aleph Alpha | 1,048,576 tokens | Apache 2.0 |
| 22 Sep 2026 | MiMo-V2.6-Flash (309B total, 15B active) | Xiaomi | 1M tokens | MIT |
| 22 Sep 2026 | MiMo-V2.6-Pro (1.02T total, 42B active) | Xiaomi | 1M tokens | MIT |
| 20 Sep 2026 | Qwen-Image-2.1 (7B generation component) | Alibaba (Qwen team) | not stated | Qwen Research License, non-commercial use only |
| 11 Aug 2026 | Nemotron 3.5 Lightning (30B total, 3B active) | NVIDIA | Up to 1M tokens | OpenMDW 1.1 |
How to choose
Start with the licence, then the size. A permissive licence lets you fine-tune and sell a product on the model; custom licences can exclude regions, cap company revenue or forbid commercial use. Total and active parameter counts in the model names show the hardware you need: mixture-of-experts models only use their active parameters per token but still have to fit all weights in memory. Prices for hosted versions are in the AI model API pricing guide, and new releases are covered in open-source AI news.
Questions people ask
- Which open-weight AI models can be used commercially without extra conditions?
- Models released under MIT or Apache 2.0 in our tracker: Granite Switch 4.1.1 (8B and 30B), Ling-3.1-flash-1019 (560B total, 25B active), Unlimited OCR 2 (about 3.3B), DeepSeek-V4.1-Flash-1017, Beam 1.0 (501B total, 23B active), Kolibri-1 (78B total, 3.46B active), MiMo-V2.6-Flash (309B total, 15B active), MiMo-V2.6-Pro (1.02T total, 42B active). Always check the model card before shipping, since terms can change.
- Which open-weight models have restrictions?
- Licences with limits such as non-commercial use, revenue caps or excluded regions: ERNIE 5.5 Flash Base (280B total, 14B active), LFM3 (six models of about 2B parameters), Qwen-Image-2.1 (7B generation component).
- What is the difference between open weights and open source?
- Open weights means the trained parameters can be downloaded and run. It does not always include training data or code, and the licence decides what you may do with the model.