Open-weight AI models compared
Open-weight models: origin, linked licences and hardware references. Sources checked 1 October 2026.
Europe
Published models from Mistral, Tilde and the EuroLLM consortium. Mistral Medium 3.5 has a revenue condition.
Model | Released | Size | Licence | What it means for you | Hardware reference |
|---|---|---|---|---|---|
| Mistral Small 4 Mistral AI (FR) | Mar 2026 | 119B MoE, 6B active | Apache 2.0 | Commercial use; licence duties apply | 4× H100 (BF16) |
| Mistral Large 3 Mistral AI (FR) | Dec 2025 | 675B MoE, 41B active | Apache 2.0 | Commercial use; licence duties apply | 8× H200 (FP8) |
| Mistral Medium 3.5 Mistral AI (FR) | Apr 2026 | 128B dense | Modified MIT | Commercial licence above $20M/month | 2× H200 (FP8) |
| EuroLLM-22B EU research consortium | Dec 2025 | 22B dense | Apache 2.0 | Commercial use; licence duties apply | 1 GPU |
| TildeOpen-30B Tilde (LV) | 2025 | 30B dense | CC BY 4.0 | Attribution required | 1 GPU |
United States
Strong permissive options from OpenAI, Google and IBM. Meta's Llama licence limits EU companies.
Model | Released | Size | Licence | What it means for you | Hardware reference |
|---|---|---|---|---|---|
| gpt-oss-120b OpenAI | Aug 2025 | 117B MoE | Apache 2.0 | Commercial use; licence duties apply | 1× 80 GB (MXFP4) |
| Gemma 4 Google | Apr 2026 | Up to 31B | Apache 2.0 | Commercial use; licence duties apply | 1 GPU |
| Nemotron 3 Ultra NVIDIA | Jun 2026 | 550B MoE, 55B active | OpenMDW 1.1 | Custom terms; review before use | 8× H200 (BF16) |
| Llama 4 Maverick Meta | Apr 2025 | 400B MoE, 17B active | Llama 4 Community | EU multimodal rights excluded | 8× H100 |
| Granite 4.2 IBM | Aug 2026 | 3B to 30B | Apache 2.0 | Commercial use; licence duties apply | 1 GPU |
China
Large open-weight options. Custom licences need review; model origin is a policy question.
Model | Released | Size | Licence | What it means for you | Hardware reference |
|---|---|---|---|---|---|
| MiMo-V2.6-Pro Xiaomi | Sep 2026 | 1.02T MoE, 42B active | MIT | Commercial use; licence duties apply | 8× B200 |
| DeepSeek V4 Pro DeepSeek | Apr 2026 | 1.6T MoE, 49B active | MIT | Commercial use; licence duties apply | 8× H200 or 8× B300 |
| Kimi K3 Moonshot AI | Jul 2026 | 2.8T MoE, 104B active | Kimi K3 License | Service licence above $20M/year | 8× B200 minimum |
| Qwen3.8 (flagship) Alibaba | Aug 2026 | 2.4T MoE, 95B active | Qwen3.8-Max License | Internal use exempt; service terms apply | Multi-node |
| GLM-5.3 Z.ai | Aug 2026 | About 750B | GLM-5.3 License | Service review above $10B/year | 8× H200 |
Licence legend
- Permissive Commercial use allowed; attribution and other licence duties still apply.
- Conditional Custom licence. Review the terms before use.
- Restricted Excludes large or EU-based companies from some rights.
Sources: linked publisher model cards and licence texts, checked 1 October 2026. The CSV records each model revision and hardware basis. Hardware is a planning reference, not a tested production minimum. Precision, context length and concurrent requests change memory needs. Licence summaries are not legal advice.
Download the October 2026 Model Index (CSV). Cite: Lindstead (2026), Open-weight AI Model Index, 1 October 2026. CC BY 4.0 applies to the index; model licences remain separate. Research questions and corrections.
From shortlist to running system
Updated . A model that fits the licence and the hardware budget is the first filter. Once a model is chosen, AI deployment on infrastructure you control takes it into production. Before that, the self-hosting break-even point shows when own hardware pays off, and European cloud GPU prices show what renting the reference hardware costs.
For large companies, a licence review checks the terms for the proposed use, and monthly model stewardship tracks new releases and licence changes. Model origin is a board policy question, discussed in the 2026 European briefing.
Choose a model for your own workload.
Lindstead tests a model shortlist on your own documents and checks every licence.