Open-weight LLM licenses for commercial use: Qwen, DeepSeek, Llama, Muse, Kimi, Mistral and 11 more compared
As of 7 November 2026, the open-weight LLMs you can use commercially with no conditions beyond keeping the notice are those under MIT, Apache 2.0 or OpenMDW-1.1: DeepSeek-V4.1-Flash, GLM-5.3-Flash, MiMo-V2.6, Qwen3.8-27B, gpt-oss, Gemma 4 and NVIDIA's Nemotron 3.5. Kimi K3.1 Pro, Meta's Muse Spark Open, Qwen3.8-Flash-Next and Mistral Large 4 add Model-as-a-Service clauses, MiniMax and Liquid add revenue limits, and Baidu's ERNIE 5.5 Flash Base excludes the EU, UK and South Korea outright. I read all 17 license texts below on 7 November 2026. This is a reading of the texts, not legal advice.
- No conditions (notice only): MIT, Apache 2.0, OpenMDW-1.1. DeepSeek, Qwen (27B), OpenAI, Google, NVIDIA, Z.ai (Flash), Xiaomi and Thinking Machines all have a current model under one of these.
- Resale clauses: Kimi K3.1 Pro and Muse Spark Open at US$20 million revenue over 12 months, Mistral Large 4 and Qwen Community 1.0 at any revenue, GLM-5.3 at US$10 billion.
- Revenue caps on all commercial use: LFM Open License (under US$10 million a year), Mistral Modified MIT (not above US$20 million a month).
- Geographic limits: ERNIE 5.5 Flash Base (no EU, UK or South Korea); Llama 4 multimodal models (no EU-based licensees).
| Family: current model | License | Commercial use | Thresholds | Attribution | Use policy | Outputs and derivatives | Geography |
|---|---|---|---|---|---|---|---|
| Qwen: Qwen3.8-27B | Apache 2.0 | Yes | None | Notice | None | No terms | None |
| Qwen: Qwen3.8-Flash-Next | Qwen Community License 1.0 | Yes, except MaaS and AI coding or office assistants | MaaS or AI Work Assistant: separate license, no revenue floor; display at 100M MAU or US$20M monthly revenue | Model name in the UI above the threshold | Comply with law | No terms | None |
| DeepSeek: V4.1-Flash | MIT | Yes | None | Notice | None | No terms | None |
| OpenAI: gpt-oss-120b | Apache 2.0 plus usage policy | Yes | None | Notice | One sentence: comply with applicable law | No terms | None |
| Meta: Muse Spark Open | Muse Model License (15 Oct 2026), gated | Yes | MaaS above US$20M revenue in 12 months: notify Meta, separate agreement | "Built with Muse Spark Open" on every product using it | Exhibit A usage policy (military, weapons, deception, under-18s) | Models trained on outputs are Derivatives | None |
| Meta: Llama 4 (for comparison) | Llama 4 Community License (5 Apr 2025) | Yes, below the MAU gate | Over 700M MAU on release date: ask Meta | "Built with Llama"; "Llama" prefix on derived model names | Llama 4 AUP | Output-trained models count, need the prefix | Multimodal models: no rights for EU-based licensees |
| Google: Gemma 4 | Apache 2.0 (page updated 1 Apr 2026) | Yes | None | Notice | None attached | No terms | None |
| Mistral: Large 4 | Mistral Ecosystem Open-Weight License | Yes, except MaaS | Any MaaS business: separate commercial license | Notice | None | No terms | None |
| Mistral: Medium 3.5, Devstral 2 | Modified MIT | Only below the cap | No rights if company revenue exceeded US$20M in the previous month | Notice | None | Cap covers derivatives | None |
| NVIDIA: Nemotron 3.5 | OpenMDW-1.1 | Yes | None | Keep the agreement and origin notices | None | Explicitly no restrictions on outputs | None |
| NVIDIA: older Nemotron (e.g. Nano 9B v2) | NVIDIA Open Model License (24 Oct 2025) | Yes | None | NOTICE line | NVIDIA Trustworthy AI terms | Outputs are not derivatives | None; rights end if you disable a safety guardrail |
| Moonshot: Kimi K3.1 Pro | Kimi K3.1-Pro License | Yes | MaaS above US$20M in 12 months: separate agreement; display at 100M MAU or US$20M monthly revenue | "Kimi K3.1-Pro" in the UI above the threshold | Comply with law | No terms | None |
| Z.ai: GLM-5.3 | GLM-5.3 License | Yes | MaaS above US$10B in 12 months: Z.ai security review | Notice | None | No terms | None |
| Z.ai: GLM-5.3-Flash | MIT | Yes | None | Notice | None | No terms | None |
| MiniMax: M3.1-Flash | MiniMax Community License | With notice; base grant reads "non-commercial" | Over US$20M yearly revenue: prior written authorization; below, a one-time email notice | "Built with MiniMax M3.1-Flash" | Prohibited uses, including any military purpose | Commercial derivatives covered | None |
| StepFun: Step-5-Preview | StepFun Model License 1.0 | Yes, except device makers | Any hardware maker shipping it as a device feature, or MaaS above US$20M in 12 months: separate license | "Powered by StepFun AI" at 100M users or US$20M monthly revenue | None separate | Models trained materially on outputs are derivatives | None |
| Baidu: ERNIE-5.5-Flash-Base | ERNIE 5.5 Flash Base License v1.0 (5 Nov 2026) | Only inside the Territory | Above US$50M annual revenue of related products: prior authorization | "Powered by ERNIE" plus a one-time notice | Exhibit A (military, high-stakes automated decisions, AI-content labelling) | Outputs may not improve any other model | Excludes EU, UK, South Korea |
| Liquid AI: LFM3 | LFM Open License v1.0 | Only below the cap | Commercial use only under US$10M annual revenue (entity plus affiliates) | Notice | None | Apache-style | None |
| Thinking Machines: Inkling | Apache 2.0 plus Model AUP (15 Jul 2026) | Yes, within the AUP | None | Notice | Separate AUP claims to bind anyone who downloads the weights | No terms | None |
Every license text read on 7 November 2026 from the linked file. The Muse Model License comes from the text Meta shows on the gated Hugging Face page; the LICENSE.md file itself needs access approval, so I could not confirm the two are identical. MiMo-V2.6 (Xiaomi), Ling-3.1-flash (inclusionAI) and DeepSeek use MIT; Ripple-65B-Preview and Arete-27B use Apache 2.0. "MaaS" is Model as a Service. MAU is monthly active users.
Which licenses allow commercial use with no conditions?
MIT, Apache 2.0 and OpenMDW-1.1 do, and in November 2026 they cover most of the strongest open models that fit modest hardware. DeepSeek's LICENSE file is the plain MIT text, and the model card says it covers both the repository and the weights. Qwen3.8-27B ships the standard Apache 2.0 text. Gemma 4 moved to Apache 2.0: its card links a Gemma 4 license page that resolves to Google's plain Apache 2.0 page, and the Gemma Terms of Use now send Gemma 4 users to that license instead. gpt-oss adds a 201-byte usage policy that asks you to follow applicable law and nothing more.
OpenMDW-1.1, which NVIDIA uses for every Nemotron 3.5 model, is the least restrictive of all. The agreement grants rights under copyright, patent, database and trade-secret law, asks only that you keep the agreement and notices when you redistribute, and states that it imposes no restrictions on outputs. Rights end only if you sue claiming the model infringes your patents or copyright. Compare NVIDIA's older Open Model License, still on models such as Nemotron Nano 9B v2: it terminates your rights automatically if you bypass or disable a safety guardrail without a substitute.
On the Open Source Initiative's definition, these are the only ones that qualify. The Open Source AI Definition 1.0 requires the freedom to use the system for any purpose without asking permission, which every revenue or resale clause below breaks.
What do the Model-as-a-Service clauses mean for you?
They mostly bite if you resell the model itself to third parties. Internal use is exempt under every license that has such a clause, and Kimi, GLM and Muse also exempt model features inside your own product; Qwen and StepFun add product categories that need a license regardless, as below. The definitions are close to identical across Kimi, GLM, Qwen and Mistral: giving a third party access to inference or fine-tuning, typically by API, with "meaningful control over the inputs, parameters, or training data". Kimi K3.1 Pro and GLM-5.3 explicitly exclude end-user products where the model is embedded in specific features. Qwen's Community License does not carve out embedded features as explicitly, but it exempts internal use that exposes nothing to third parties.
The revenue triggers run from zero to US$10 billion:
| License | MaaS trigger | What happens | Revenue counted |
|---|---|---|---|
| Mistral Ecosystem (Large 4) | Any MaaS business | Separate commercial license | n/a |
| Qwen Community 1.0 (Flash-Next) | Any MaaS or AI Work Assistant business | Separate license before commercial use | n/a |
| Kimi K3.1-Pro | Over US$20M in any 12 months | Separate agreement | Licensee plus affiliates, all revenue |
| Muse Model License | Over US$20M in any 12 months | Notify Meta, separate agreement | Gross revenue, licensee plus affiliates |
| StepFun Model License 1.0 | Over US$20M in any 12 months | Separate license | Combined revenue from all sources |
| GLM-5.3 License | Over US$10B in any 12 months | Z.ai security review | Licensee plus affiliates |
From the license texts linked in the main table, read 7 November 2026.
Note what the Qwen clause adds: "AI Work Assistant" means an independent product mainly for AI-assisted coding or office work, and the license names Qoder and QwenWork as examples. If you build a coding assistant product on Qwen3.8-Flash-Next, you need a separate license from Qwen regardless of revenue. Qwen3.8-27B, under Apache 2.0, has no such clause.
The display thresholds are easy to misread too. Kimi, Qwen and StepFun require the model name or "Powered by StepFun AI" at 100 million monthly users or US$20 million revenue per month, which by my arithmetic is US$240 million a year. MiniMax's US$20 million threshold for prior authorization is per year, twelve times lower, and its "Built with MiniMax M3.1-Flash" notice applies to any commercial use at all.
Which licenses restrict use in the EU?
Two do, in very different ways. Baidu's ERNIE 5.5 Flash Base license (version 1.0, dated 5 November 2026) defines its Territory as everywhere except the European Union, the United Kingdom and South Korea, and says the model, derivatives and outputs must not be used, distributed or displayed outside it. For a company in the EU, that means no rights at all.
Meta's Llama 4 acceptable use policy withholds the license grant for multimodal Llama 4 models from individuals domiciled in the EU and companies with their principal place of business there; end users of a product built on them are exempt. Meta's newer Muse Model License drops this clause, along with the 700 million MAU gate and the "Llama" naming rule, and names Meta Platforms Ireland as the licensor for the European region.
None of the other licenses mentions geography. Regulation is a separate question: what the EU AI Act asks of a company that self-hosts an open model is covered in the planned article on the EU AI Act and open-weight models, and GPUwerk's EU compliance page covers GDPR on dedicated hardware.
Can you train another model on the outputs?
Under MIT, Apache 2.0 and OpenMDW-1.1, nothing in the license prevents it. Four licenses do reach outputs:
- ERNIE 5.5 Flash Base: section 5.5 forbids using the model, derivatives or outputs to improve any other AI model.
- Muse Model License: any model created, trained, fine-tuned or distilled using the model or its outputs is a Derivative, and every product using a Derivative must show "Built with Muse Spark Open". Meta claims no ownership of outputs.
- StepFun Model License 1.0: a model trained to a material extent on its outputs is a Derivative Work, so the device-maker and MaaS clauses follow it.
- Llama 4 Community License: a model built with Llama materials or outputs and then distributed must have "Llama" at the start of its name.
NVIDIA's Open Model License goes the other way and says outputs are not derivative models. Nous Research's Arete-27B is a working example of the permissive path: it fine-tunes Apache 2.0 Qwen3.8-27B on data generated by Inkling (Apache 2.0) and ships under Apache 2.0.
Which use policies bind the weights?
Muse Spark Open, MiniMax-M3.1-Flash and ERNIE 5.5 Flash Base come with a use policy that is part of the license, as Llama 4 does, and Inkling ships an Apache license with a separate policy that claims to bind downloaders anyway.
- Muse Spark Open: Exhibit A, dated 15 October 2026, follows the Llama 4 policy (military and weapons, critical infrastructure, deception, presenting outputs as human-written) and adds that the model is not intended for people under 18.
- MiniMax-M3.1-Flash: an appendix bans any military purpose, harming minors, harmful disinformation and discrimination.
- ERNIE 5.5 Flash Base: Exhibit A bans military use and high-stakes automated decisions, and requires labelling of machine-generated content you publish.
- Inkling: the weights are Apache 2.0, but Thinking Machines' Model Acceptable Use Policy (15 July 2026) says that downloading or using the weights binds you to it. Whether a separate policy can narrow an Apache grant is a question for a lawyer; I would treat it as binding.
Gemma 4 is the reverse case: the old Gemma Prohibited Use Policy is incorporated by the Gemma Terms of Use, which no longer cover Gemma 4, and the Gemma 4 cards do not link it.
Which license should a company pick?
If you are an EU company that may one day sell access to a model, pick MIT, Apache 2.0 or OpenMDW and stop there. That still leaves strong options at every size: Qwen3.8-27B for one GPU, Nemotron 3.5 Super VL (OpenMDW) or gpt-oss-120b for one 128 GB machine, GLM-5.3-Flash or MiMo-V2.6-Flash (both MIT) for two, and DeepSeek-V4.1-Flash for a GPU server. Where they rank against each other is in the open-weight leaderboard snapshot.
Kimi K3.1 Pro and Muse Spark Open are reasonable for internal use or a product feature; read their MaaS and display clauses before you expose an API. Avoid ERNIE 5.5 Flash Base in the EU, LFM3 above US$10 million revenue, and Qwen3.8-Flash-Next for a coding-assistant product. Mistral's Modified MIT models are off limits to any company that has passed US$20 million in monthly revenue.
Licenses change between releases, often within a family: GLM-5.3 and GLM-5.3-Flash differ, as do Qwen3.8-27B and Qwen3.8-Flash-Next. Pin the license text with the model revision you deploy, the same way you pin the weights. To try the permissive models on your own data before any of that matters, a Spark rents at $0.79/hour (pricing), and the models guide lists what fits in 128 GB.
FAQ
Which open-weight LLMs can I use commercially without conditions?
Models under MIT, Apache 2.0 or OpenMDW-1.1 carry no revenue, user-count or use-case conditions beyond keeping the license notice. As of 7 November 2026 that includes DeepSeek-V4.1-Flash (MIT), GLM-5.3-Flash (MIT), MiMo-V2.6 (MIT), Qwen3.8-27B, gpt-oss-120b and Gemma 4 (Apache 2.0), and NVIDIA's Nemotron 3.5 models (OpenMDW-1.1), per the license files on Hugging Face.
Is the Llama license open source?
No, by the Open Source Initiative's standard. Its Open Source AI Definition requires freedom to use a system for any purpose without asking permission, and the OSI said in 2023 that Llama's license fails that test. The Llama 4 Community License still limits companies with over 700 million monthly active users and, through its acceptable use policy, withholds rights for multimodal Llama 4 models from EU-based companies.
Can a company in the EU use Llama 4 or Muse Spark Open?
For Llama 4's multimodal models, Meta's acceptable use policy does not grant the license rights to individuals domiciled in the EU or companies with their principal place of business there, though end users of a product are exempt. Muse Spark Open's license, dated 15 October 2026, has no such EU clause. Baidu's ERNIE 5.5 Flash Base license excludes the EU, UK and South Korea entirely.
What does a Model-as-a-Service clause mean in an LLM license?
It targets companies that resell inference or fine-tuning of the model to third parties, usually through an API. Kimi K3.1 Pro and Muse Spark Open require a separate agreement once such a business passes US$20 million in revenue over 12 months, Mistral Large 4 requires one at any revenue, and Qwen's Community License 1.0 also covers AI coding and office assistants. Internal use is exempt in all of them.
Can I train my own model on another model's outputs?
Under MIT, Apache 2.0 and OpenMDW-1.1, nothing in the license stops it. Baidu's ERNIE 5.5 Flash Base license forbids using outputs to improve any other model. StepFun's license and Meta's Muse license treat models trained on outputs as derivatives, so their terms follow, including Muse's 'Built with Muse Spark Open' display duty. The Llama 4 license requires 'Llama' at the start of such a model's name.
Sources
Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.
- Qwen Community License 1.0 (Qwen3.8-Flash-Next LICENSE), Qwen / Hugging Face
- Qwen3.8-27B LICENSE (Apache 2.0), Qwen / Hugging Face
- DeepSeek-V4.1-Flash LICENSE (MIT), DeepSeek / Hugging Face
- gpt-oss-120b USAGE_POLICY, OpenAI / Hugging Face
- Muse-Spark-Open-816B-A42B model card and license, Meta / Hugging Face, 2026-10-15
- Llama 4 Community License Agreement, Meta, 2025-04-05
- Llama 4 Acceptable Use Policy, Meta
- Gemma 4 license (Apache 2.0), Google AI for Developers, 2026-04-01
- Mistral Ecosystem Open-Weight License (Mistral Large 4 LICENSE), Mistral AI / Hugging Face
- Mistral Modified MIT License (Mistral-Medium-3.5-128B LICENSE), Mistral AI / Hugging Face
- OpenMDW License Agreement 1.1, OpenMDW / Linux Foundation
- NVIDIA Open Model License Agreement, NVIDIA, 2025-10-24
- Kimi K3.1-Pro License, Moonshot AI / Hugging Face
- GLM-5.3 License, Z.ai / Hugging Face
- MiniMax Community License (MiniMax-M3.1-Flash), MiniMax / Hugging Face
- StepFun Model License 1.0 (Step-5-Preview), StepFun / Hugging Face
- ERNIE 5.5 Flash Base Model License Agreement v1.0, Baidu / Hugging Face, 2026-11-05
- LFM Open License v1.0 (LFM3-1.7B), Liquid AI / Hugging Face
- Model Acceptable Use Policy, Thinking Machines Lab, 2026-07-15
- The Open Source AI Definition 1.0, Open Source Initiative