License selection materially shapes the production deployability of every open-weight model. The 2026 landscape includes Apache 2.0 and MIT (most permissive), Llama Community Licence and Tongyi Qianwen (permissive with scale and competitive-use restrictions), Gemma Terms of Use, OpenRAIL-M (use-case restrictions), CC-BY-NC (research-only), and dozens of vendor-specific custom licences. This page consolidates the license landscape, adoption share, and procurement guidance.
Key Findings
- Apache 2.0 is the most-adopted open-weight license in 2026 with approximately 38 percent of new releases on Hugging Face. The unrestricted commercial use and explicit patent grant remove most procurement friction.
- MIT is the second-most-adopted at approximately 18 percent, with the broadly-deployed Phi family, Granite family, and several Chinese lab releases (InternVL, OLMo) using MIT.
- Llama Community Licence (3.x and 4) covers approximately 14 percent of new releases by count and a much larger share by downloads. The licence permits commercial use but has a 700M monthly active user threshold above which an explicit Meta licence is required.
- Tongyi Qianwen Licence (Qwen family) is used in approximately 6 percent of new releases. The licence permits commercial use but has restrictions on competitive AI service use and on scale (100M MAU threshold for explicit licence).
- CC-BY-NC remains common for research-focused releases (~9 percent of new releases) but restricts commercial deployment without separate negotiation.
Open-Weight License Adoption (May 2026)
| Licence | Share of New Releases | Commercial Use | Restrictions |
|---|---|---|---|
| Apache 2.0 | ~38% | Unrestricted | None |
| MIT | ~18% | Unrestricted | None |
| Llama Community Licence | ~14% | Permitted | 700M MAU threshold, name-attribution rules |
| Tongyi Qianwen Licence | ~6% | Permitted | 100M MAU threshold, restrictions on competitive AI services |
| Gemma Terms of Use | ~5% | Permitted | Prohibited use policy, requires user policy acceptance |
| OpenRAIL-M | ~3% | Permitted with use restrictions | Use-case restrictions (prohibited use list) |
| CC-BY-NC variants | ~9% | Research only by default | Commercial use requires separate licence |
| Custom commercial | ~5% | Permitted with negotiation | Vendor-specific (e.g., Cohere) |
| BSD / GPL / AGPL | ~2% | Permitted (GPL with copyleft) | Source-distribution requirements for GPL |
Major Models by License
| Licence | Notable Models |
|---|---|
| Apache 2.0 | Qwen2.5 / Qwen3 (most variants), Mistral 7B / Small 3, FLUX.1 Schnell, Granite 3.x, OLMo 2, IBM Granite, Wan 2.1, Mochi-1, Allegro |
| MIT | Phi-3 / Phi-4 family, BGE family, InternVL3, DeepSeek-R1 family, OLMo, RankZephyr, Open-Sora 2.0 |
| Llama Community | Llama 3.1 / 3.2 / 3.3 / Llama 4 family, Llama-Nemotron family, Hermes 3 / 4 (Llama backbones) |
| Tongyi Qianwen | Qwen2.5-VL-72B, Qwen3-Embedding, Qwen3-Reranker, Command R+ analogues |
| Gemma Terms | Gemma 2, Gemma 3, PaliGemma, ShieldGemma |
| OpenRAIL-M / variants | SDXL, LTX-Video, Stable Diffusion family |
| CC-BY-NC | NV-Embed-v2, NVLM, ChatTTS, Linq-Embed-Mistral, SFR-Embedding-Mistral |
Procurement Decision Framework
| Procurement Scenario | License Recommendation |
|---|---|
| Standard commercial product deployment | Apache 2.0 or MIT (zero negotiation friction) |
| Startup with under 700M MAU | Apache 2.0, MIT, or Llama Community Licence (no immediate friction) |
| Hyperscaler / large platform (over 700M MAU) | Apache 2.0 or MIT only; Llama Community Licence requires explicit Meta agreement |
| Competitive AI product | Apache 2.0 or MIT only; Tongyi Qianwen restricts competitive AI service use |
| Research and academic | Any licence (CC-BY-NC and OpenRAIL-M acceptable for research) |
| Regulated industry (financial, healthcare) | Apache 2.0 preferred for legal simplicity; MIT acceptable |
| Government / defence | Apache 2.0 preferred; some agencies have specific licence restrictions |
Strategic Context
Three patterns shape the 2026 licence landscape. First, Apache 2.0 is the procurement-friendly default and is increasingly the chosen licence by labs that prioritise adoption (Qwen for 7B variants, Granite, OLMo). Second, the Llama Community Licence is the most consequential restricted licence: the 700M MAU threshold materially complicates hyperscaler-scale deployment but leaves the vast majority of commercial use unaffected. Third, CC-BY-NC remains common for research releases but introduces negotiation friction; production teams typically substitute for Apache or MIT equivalents.
Brand Visibility Implications
License complexity is a major procurement decision factor. AI assistant queries about "Llama 3 commercial use", "Tongyi Qianwen license", "open-source LLM commercial", and similar terms drive procurement-research traffic. Brands selling AI counsel, AI procurement advisory, and license-compliance tooling face strong AI-mediated discovery surface for this category.
Methodology
License adoption shares sampled from new Hugging Face model releases by parent organisation through Q1-Q2 2026. License terms compiled from primary licence text disclosures. Updated quarterly.
How Presenc AI Helps
Presenc AI monitors brand visibility on AI license queries across ChatGPT, Claude, Gemini, and Perplexity. For AI counsel brands, AI procurement advisors, and license-compliance tooling vendors, the platform identifies the prompts driving research-traffic patterns and the gaps where new content unlocks share of voice.