Scoreboard / Models
How open is it, really?
“Open” covers a lot of ground. Some models come with no strings. Others are free until you get big. We read the licences so you don't have to.
| Grade | Model | Licence | What the terms say |
|---|---|---|---|
| A | DeepSeek V4.1 FlashDeepSeek · China | MIT | No conditions beyond keeping the licence notice. |
| A | GLM-5.3 FlashZ.ai · China | MIT | No conditions beyond keeping the licence notice. |
| A | Qwen3.8 27BAlibaba · China | Apache 2.0 | Standard permissive terms. |
| A | Muse Glimmer 30BMeta · U.S. | Apache 2.0 | Standard permissive terms. |
| A | gpt-oss-120bOpenAI · U.S. | Apache 2.0 | Standard permissive terms, plus a short usage policy. |
| A | Gemma 4Google · U.S. | Apache 2.0 | Standard permissive terms. |
| A | Nemotron 3 UltraNvidia · U.S. | OpenMDW 1.1 | Permissive. Keep the licence and origin notices when you redistribute. |
| B | Kimi K3.1 ProMoonshot AI · China | Kimi K3.1 Pro License | MIT-style, with extra conditions for model-hosting businesses above $20 million in revenue and for products with 100 million monthly users. |
| B | Muse Spark OpenMeta · U.S. | Muse Model License | Free commercial use. Requires a visible “Built with Muse Spark Open” credit, and a separate agreement from companies over $20 million in revenue that resell model access. |
| B | GLM-5.3Z.ai · China | GLM-5.3 License | Adds a security review for model-service providers above $10 billion in revenue. |
| B | Qwen3.8 Max (2.4T)Alibaba · China | Qwen3.8-Max License | Adds display and licensing conditions at scale. |
| F | Claude Opus 5.5Anthropic · U.S. | Proprietary | No downloadable weights. Access through Anthropic's services. |
| F | GPT-5.6 SolOpenAI · U.S. | Proprietary | No downloadable weights. Access through OpenAI's services. |
| F | Gemini 4 ArgonGoogle · U.S. | Proprietary | No downloadable weights. Access through Google's services. |
| F | Muse WonderMeta · U.S. | Proprietary | No downloadable weights. Access through Meta's API. |
Why nobody gets an A+
Open weights are not the same as open source. You can run and change these models, but you can't rebuild them. Mozilla reviewed 16 notable open releases in September and found that none publishes the full recipe for its training data. Until one does, A is the top grade.
Sources
- 9 Best Open Source LLMs to Self-Host, Maxim, Nov 2026
- What Are Open Weight Models? 8 Worth Running in 2026, Respan, Nov 2026
- Open-Source LLMs 2026: Which Are Actually Open-Licensed, Codersera, Oct 2026
- Best Open-Weight LLMs 2026: Local Models Compared, Sysart, Oct 2026
- Kimi K3.1 Pro: Benchmarks, API Pricing, Open Weights, Morph, Nov 2026
- Meta unveils Muse Spark Open, the new top U.S. open weights model, VentureBeat, Oct 2026
- Muse Spark Open tested: specs, licence, hardware and cost, AI Wiz, Oct 2026
- Meta's 816B Muse Has a $20M API Line; 30B Glimmer Has None, FourWeekMBA, Oct 2026
- What Are Open-Weight Models? The Latest Releases, deepset, Nov 2026
Licences checked Nov 21, 2026. Grades cover licence terms only. They say nothing about how good or how safe a model is. Read the licence yourself before you ship.