Truly open weights, and the licenses that only look open
"Open weight" covers everything from Apache-2.0 — self-host it, ship it in a product, no strings — to licenses that forbid commercial use or bolt on acceptable-use clauses. Across 165 models here, 50% are permissively licensed (Apache / MIT), 25% are open but conditional, and 25% are proprietary — weights closed, API only. A license belongs to the model, identical whichever provider serves it, so each figure below is one host per model (the cheapest).
Cheapest genuinely open model: mistral-nemo ships Apache 2.0 weights — commercial use and self-hosting, no clause — at $0.038/1M blended (Io Net).
Permissive means the license itself won't stop you shipping it. It says nothing about running cost — see output vs input and best value per dollar for that.
How open is the field?
Share of models in each license tier.
Licenses in the index
| License | Tier | Models | Reference |
|---|---|---|---|
| Apache 2.0 | permissive | 47 | terms → |
| Proprietary | proprietary | 37 | terms → |
| MIT | permissive | 36 | terms → |
| Llama Community License | conditional | 10 | terms → |
| Modified MIT | conditional | 9 | terms → |
| CC-BY-NC 4.0 | conditional | 5 | terms → |
| NVIDIA Open Model License | conditional | 5 | terms → |
| Gemma Terms of Use | conditional | 4 | terms → |
| Proprietary — Mistral « Premier » | proprietary | 4 | terms → |
| GLM-5.3 License | conditional | 1 | terms → |
| MiniMax Community License | conditional | 1 | terms → |
| MiniMax Model License | conditional | 1 | terms → |
| MiniMax Non-Commercial License | conditional | 1 | terms → |
| Muse Model License | conditional | 1 | terms → |
| OpenMDW-1.1 | permissive | 1 | terms → |
| Qwen License Agreement | conditional | 1 | terms → |
| Qwen3.8-Max License | conditional | 1 | terms → |
Cheapest truly-open models
Permissively licensed (Apache / MIT) — self-host, run commercially, no clause — ranked by price at their cheapest host.
| Model | License | Cheapest host | Blended /1M |
|---|---|---|---|
| mistral-nemo | Apache 2.0 | Io Net | $0.038 |
| gpt-oss-20b | Apache 2.0 | AkashML | $0.12 |
| gpt-oss-120b | Apache 2.0 | Venice | $0.18 |
| glm-5.3-flash | MIT | StreamLake | $0.189 |
| ministral-3b | Apache 2.0 | Mistral AI | $0.2 |
| deepseek-v4-flash | MIT | Relace | $0.21 |
| deepseek-v4-flash-0423 | MIT | StreamLake | $0.21 |
| ling-3.0-flash | MIT | Novita AI | $0.24 |
| ling-3.0-flash-vl | MIT | DeepInfra | $0.24 |
| qwen3-30b-a3b | Apache 2.0 | StreamLake | $0.241 |
| qwen3.5-9b | Apache 2.0 | SiliconFlow | $0.25 |
| granite-4.2-8b | Apache 2.0 | CoreWeave | $0.25 |
| gemma-4-26b-a4b | Apache 2.0 | Reka | $0.26 |
| mistral-small-3.2-24b | Apache 2.0 | DeepInfra | $0.275 |
| ling-3.0-flash-fin | MIT | Novita AI | $0.295 |
| ministral-8b | Apache 2.0 | Mistral AI | $0.3 |
| ling-3.1-flash | MIT | DeepInfra | $0.3 |
| qwen3-14b | Apache 2.0 | NextBit | $0.32 |
| deepseek-v3.2-exp | MIT | AtlasCloud | $0.336 |
| qwen3-coder-30b-a3b | Apache 2.0 | Novita AI | $0.34 |
Prices per 1M tokens (USD), each model's cheapest published rate. Each license links to the terms it was read from — the LICENSE file in the maker's own weights repository where we have verified it model by model, the family's reference license otherwise. Verified 2026-10-23.
FAQ
What does 'open weight' actually mean for an LLM?
It ranges widely. Across 165 models here, 50% are permissively licensed (Apache/MIT — self-host and ship commercially, no strings), 25% are open but conditional (non-commercial, acceptable-use, attribution or large-scale commercial clauses), and 25% are proprietary — weights closed, API only.
What's the cheapest truly-open model I can self-host and ship?
mistral-nemo ships Apache 2.0 weights — commercial use and self-hosting with no clause — at $0.038 per 1M tokens blended via Io Net, the cheapest genuinely open option in the index.
Is a permissive license enough to use a model in production?
The license only settles whether you're allowed to ship it — permissive means it won't stop you. It says nothing about running cost or quality; check the value ranking and the model's per-token price separately.