Open-Weight LLM Licences and Self-Hosting Footprint (2026)
As of 2026. Licence families are high-confidence for major publishers but exact terms vary per model and change with releases - verify the specific model card before relying on any cell.
| Family | Publisher (HF org) | 30d downloads (top-8) | Licence family | Commercial use | Typical self-host footprint |
|---|---|---|---|---|---|
| Qwen | Alibaba (Qwen) |
~113M | Mostly Apache-2.0 (some sizes under Qwen License) | Yes (mostly) | ~0.5B - 72B+, dense + MoE |
| Gemma | Google (google) |
~97M | Gemma Terms of Use (custom, not OSI) | Yes, with use restrictions | ~2B - 27B |
| Llama | Meta (meta-llama) |
~29M | Llama Community License (custom) | Yes, with a large-scale (>700M MAU) restriction + naming rules | ~1B - 405B (3.x); MoE (4.x) |
| DeepSeek | DeepSeek (deepseek-ai) |
~19M | Code MIT + DeepSeek Model License; recent releases MIT-friendly | Yes | up to ~671B MoE (~37B active) |
| Phi | Microsoft (microsoft) |
~19M | MIT | Yes | small (~3.8B - 14B) |
| GLM | Zhipu (zai-org / THUDM) |
~12M | Mixed (MIT/Apache for recent GLM-4; some custom) | Mostly yes (verify) | ~9B - large MoE |
| Mistral | Mistral (mistralai) |
~10M | Apache-2.0 (open models) + Mistral Research/Commercial License (flagship) | Split: open models yes; flagship research-only unless licensed | 7B - 8x22B MoE |
| Kimi | Moonshot (moonshotai) |
~5M | Modified MIT | Yes | large MoE (K2 ~1T total / ~32B active) |
| Falcon | TII (tiiuae) |
~1.6M | Apache-2.0 (recent); earlier TII Falcon License | Yes (recent) | up to ~180B (earlier); Falcon 3 smaller |
| Command / Command-R | Cohere (CohereLabs) |
~1.2M | CC-BY-NC 4.0 (non-commercial) + acceptable-use | No - non-commercial weights (commercial via Cohere) | ~35B - 104B |
| Yi | 01.AI (01-ai) |
~0.15M | Apache-2.0 (recent) / earlier Yi License | Yes (recent) | ~6B - 34B |
The finding for Section 5
- Most open-weight families are commercially self-hostable under Apache-2.0, MIT, or a permissive custom licence - which is what makes self-hosting a realistic mainstream stance, not a research-only one.
- But the licence is a genuine decision axis, not a formality. Cohere's Command family is non-commercial (CC-BY-NC); Mistral's newest flagship models are research-only unless separately licensed; Llama and Gemma carry custom terms with real restrictions (scale caps, naming, acceptable-use). A team choosing an open model on capability alone can back into a licensing problem.
- Footprint ranges from laptop-class to data-centre-class. Phi and small Qwen/Gemma variants run on modest hardware; DeepSeek and Kimi's large MoE models are serious infrastructure. "Self-hostable" spans two orders of magnitude of cost - worth making explicit for buyers.