Which open source LLMs can you use commercially in 2026?

If you need an open source LLM for commercial use, the licences with the fewest conditions are MIT and Apache 2.0, and as of 9 October 2026 the strongest open-weight model is under one of them. Xiaomi’s MiMo-V2.6-Pro scores 46 on Artificial Analysis’s Intelligence Index, the highest of any open-weight model, and it is MIT-licensed. DeepSeek V4, DeepSeek V4.1-Flash and Z.ai’s GLM-5.3-Flash are MIT too. Qwen3.8-27B, Gemma 4, gpt-oss, Mistral Small 4 and Meta’s Muse Glimmer are Apache 2.0.

Most of the other strong models allow commercial use with conditions. GLM-5.3, Kimi K3 and the two larger Qwen3.8 models add duties for companies that sell access to models, and Qwen’s also reach AI coding and office-productivity products. Kimi and Qwen want the model’s name in your interface once a product passes 100 million monthly users or US$20 million in monthly revenue. Llama 4 requires a “Built with Llama” notice and withholds its multimodal models from companies based in the EU. Mistral Medium 3.5 gives no rights at all to a company earning more than US$20 million a month.

This guide goes through nine model families, linking each licence and setting out its thresholds, use restrictions, naming duties and output rules alongside each model’s score. It ends with default choices and a checklist for shipping a self-hosted model. It is not legal advice, so have a lawyer read any custom licence before you rely on it.

Is an open source LLM free for commercial use?

That depends on the licence attached to the weights. The Open Source Initiative’s Open Source AI Definition 1.0 asks for more than open weights. It also wants enough information about the training data for a skilled person to build a substantially equivalent system, the complete code used to train and run it, and the freedom to use it for any purpose without asking permission.¹ The labs below publish weights under a licence of their choosing, so this guide calls their models open-weight and judges each one by its licence.

Three kinds of licence cover the models here.

• MIT lets you use, copy, modify, publish, distribute, sublicense and sell the software, provided the copyright and permission notice go with all copies or substantial portions of it.² It says nothing about patents.
• Apache 2.0 gives the same freedoms plus an express patent licence from each contributor, which ends if you sue claiming the work infringes a patent. Anyone who redistributes must include the licence, mark the files they changed and carry over any NOTICE file. It grants no trademark rights.³
• Custom licences mostly start from MIT wording and add conditions, such as revenue or user thresholds, a model name you must display, an acceptable-use policy or rules for models trained on the outputs. Hugging Face tags them license:other.

The licence covers the weights you download. If you call a lab’s hosted API instead, its terms of service apply, and they are a separate document. The notice duties in MIT and Apache 2.0 attach to copies you distribute, and an app that bundles the weights distributes them.

Open-weight model licences compared, as of 9 October 2026

Each licence link goes to the licence file in the model’s Hugging Face repository. Where the repository has no public licence file, it goes to the model card or to the licence page the card names. “MaaS” means Model as a Service, which the GLM, Kimi and Qwen licences define as giving third parties inference or fine-tuning access with meaningful control over the inputs, parameters or training data. Scores are Artificial Analysis Intelligence Index v4.3.2, and an asterisk marks Artificial Analysis’s own estimate where its independent run is still to come.⁴

We read each licence file, model card and Hugging Face licence tag linked here on 9 October 2026, and took the scores from Artificial Analysis on the same day. We did not run any of the models. The guide leaves out other open-weight families, such as MiniMax and NVIDIA’s Nemotron, and the licences of training datasets.

Mistral Large 4’s score is for the API preview Mistral launched on 6 October. Mistral says it will release the weights by the end of the month, and it has not named a licence.⁵

MIT and Apache 2.0 LLM models with no thresholds

These models can go into a commercial product with no revenue line, user count or naming rule beyond the licence notice. Their sizes and scores are in the capability table further down.

DeepSeek V4 and V4.1-Flash. The cards for V4-Pro, V4-Pro-0813 and V4.1-Flash each say that the repository and the model weights are licensed under MIT, and each repository holds the MIT licence file.⁶ V4.1-Flash takes images as well as text.

Xiaomi MiMo-V2.6. Xiaomi has published two Pro checkpoints, MiMo-V2.6-Pro-RL and a later MiMo-V2.6-Pro-MOPD that reduces repeated tool calls. Both cards declare MIT in their metadata, but neither repository contains a licence file or a copyright line.⁷ Artificial Analysis’s score of 46 links to the RL weights. Because the licence lives only in the card, save the card at the revision you download.

GLM-5.3-Flash. Z.ai’s smaller GLM-5.3 model is MIT-licensed, while the larger GLM-5.3 has a custom licence.⁸

Qwen3.8-27B. The only Qwen3.8 model under Apache 2.0 is this 27B dense model. At 34 it has the highest score in Artificial Analysis’s 4B-to-40B size class, nine points ahead of the next model.⁹⁴

Gemma 4. Google’s Gemma 4 cards link to an Apache 2.0 licence page. The Gemma Terms of Use, last modified on 1 April 2026, now apply only to the models in their appendix, which lists Gemma 1 to Gemma 3n and their variants, and send Gemma 4 users to the Apache licence.¹⁰ The Gemma 4 repositories are also ungated, while Gemma 3’s require an access request.

gpt-oss. OpenAI’s two models from August 2025 are Apache 2.0, with a two-sentence usage policy whose only requirement is that you comply with applicable law.¹¹

Mistral Small 4. The card and its metadata say Apache 2.0, though the repository has no licence file.¹²

Muse Glimmer is Apache 2.0 as well, with a usage policy alongside it, and the section on Meta below covers it.

Custom licences with revenue thresholds

GLM-5.3, Kimi K3, the two larger Qwen3.8 models and Mistral’s Modified MIT models all allow commercial use, and each draws a line somewhere. The Z.ai, Moonshot and Qwen licences tie their main condition to a Model as a Service business. Z.ai and Moonshot exclude end-user products whose model capabilities sit inside specific features, and plain relaying of requests to models hosted by others. Qwen’s definition excludes only the relaying.⁸¹³⁹ The table sets out each line as the licence texts stood on 9 October 2026.

In the GLM, Kimi and Qwen revenue tests, “group” means the licensee together with its affiliates. A small product team inside a large company is measured by the whole company.

Kimi K3’s licence lifts both of its clauses for internal use, meaning use that makes neither the software, its outputs nor its capabilities available to third parties. It also lifts them for use through Moonshot’s own products or certified inference partners, and it says nothing similar about other hosts.¹³

The Qwen licences define an AI Work Assistant as an independent product designed mainly for AI-assisted coding or office productivity, and give Qoder and QwenWork as examples. Single-purpose tools such as translators fall outside it, as do assistants for other domains and assistants that are one feature of a product with a different main purpose. Under the Qwen Community License 1.0, a company selling an AI coding tool needs a separate licence before any commercial use of Qwen3.8-Flash-Next, whatever its revenue. Both Qwen licences exempt internal use from their Model as a Service and AI Work Assistant clause in the same way Kimi’s does.⁹

Mistral’s Modified MIT, used for Mistral Medium 3.5 and Devstral 2, works differently. It applies to every company, whatever it builds, and above its line it grants no rights at all, including for internal use. The test is monthly, so US$20 million a month works out at US$240 million a year. The limit also covers derivatives and fine-tunes, whether Mistral or a third party provides them.¹²

The Llama licence for commercial use, and Muse Glimmer

Meta now publishes open weights under two very different licences.

Llama 4 Scout and Maverick use the Llama 4 Community License Agreement, effective 5 April 2025. It grants a royalty-free licence for commercial use, with these conditions.¹⁴

• Companies whose products passed 700 million monthly active users in the calendar month before the release must ask Meta for a licence.
• If you distribute the models, or a product or service that contains them, you must include a copy of the agreement and prominently display “Built with Llama” on a related website, interface, blog post, about page or product documentation. Copies of the weights also need a Notice file carrying Meta’s attribution line.
• Your use must follow the Llama 4 Acceptable Use Policy, which the licence incorporates. Among other things, it bars the unlicensed practice of professions such as finance, law and medicine, the operation of critical infrastructure or heavy machinery, and presenting outputs as written by a human.
• For the multimodal models in Llama 4, the policy grants no rights to people living in the EU or to companies with their principal place of business there. End users of a product built on them are unaffected. Scout and Maverick both take image input.⁴
• If you sue Meta or anyone else claiming that the Llama materials or their outputs infringe your intellectual property, your licence ends.

The Llama 4 repositories on Hugging Face are gated, and Meta’s meta-llama organisation hasn’t published a model there since April 2025. Artificial Analysis estimates Maverick at 10 and Scout at 8.

Muse Glimmer, which Meta released in August 2026, is plain Apache 2.0, and its card says it is intended for commercial and research use.¹⁵ The repository also holds a usage policy that says it applies to your access or use of the model. Its prohibited uses largely repeat the Llama 4 policy, and it adds that the model is not intended for anyone under 18. It has no EU clause, user threshold or naming rule. Apache 2.0 itself contains no use restrictions, so ask your lawyer how the policy binds you, and follow it in the meantime, particularly in a consumer product.

Both Meta policies forbid presenting outputs as human-generated, which overlaps with the disclosure duties that took effect in the EU and California in August. Our post on what EU and California law now require of AI-generated content covers those duties.

Download Muse Glimmer from Meta’s verified meta-models organisation on Hugging Face, the source Artificial Analysis links to. A search on Hugging Face also returns copies under unverified accounts with Meta-like names.

Training other models on the outputs

Llama 4 is the only current licence in this guide with a rule on training other models. If you use the Llama materials or their outputs to create, train, fine-tune or improve an AI model that you distribute or make available, that model’s name must begin with “Llama”.¹⁴

Google’s older terms have a rule of their own. The Gemma Terms of Use, which still cover Gemma 1 to 3n, count a model as a Model Derivative if it was trained to perform like Gemma through distillation or on synthetic data that Gemma generated, and the terms apply to Model Derivatives.¹⁰ Gemma 4’s Apache 2.0 licence has no such definition.

MIT and Apache 2.0 don’t mention outputs. The GLM-5.3, Kimi K3 and Qwen licences mention them in their warranty disclaimers, and Kimi’s and Qwen’s also do so in their internal-use exemptions, but none of them limits training.

All of this concerns the weights. Outputs you collected from a lab’s hosted API fall under that API’s terms of service, which you need to read separately.

The best open source LLMs in 2026, by score and licence

The Intelligence Index v4.3.2 combines ten evaluations, including Terminal-Bench 4.0, Humanity’s Last Exam and several agent and knowledge-work tests. The table shows each model at the highest reasoning setting Artificial Analysis ran, as of 9 October 2026.⁴ For comparison, the top closed model, Claude Opus 5.5, scores 58, and our LLM API price and benchmark comparison covers the closed models.

Parameter counts come from the model cards and Mistral’s announcement, except the totals for GLM-5.3 and Llama 4, which are Hugging Face’s count of the published weights. Llama 4’s 17B active figure is part of its model names. DeepSeek V4.1-Flash activates 8B parameters per token while reading the prompt and 16B while generating. An asterisk marks an Artificial Analysis estimate.

Sticking to MIT and Apache 2.0 costs nothing at the top of the table, because MiMo-V2.6-Pro leads at 46. The best custom-licence models, GLM-5.3 and Kimi K3, sit one and two points behind it.

The open Qwen3.8-2.4T-A95B scores 40, five points below Qwen3.8-Max, the API model built on it, which adds image input, a non-thinking mode and a 1M-token context by default.⁹⁴

Scores also depend on the reasoning setting. Artificial Analysis has GLM-5.3 at 45 on its maximum setting and 34 on low, Kimi K3 at 44 and 30, and Qwen3.8-27B at 34 on xhigh and 26 on low. A self-hosted copy run at a cheap setting can land well below the table.

All nine open-weight models that score 35 or more come from Xiaomi, Z.ai, Moonshot, Alibaba and DeepSeek, which are all based in China. If your customers restrict where a model comes from, note that Mistral Large 4, at 38 as a preview, would be the first model in that group from outside China if its weights ship under a licence you can use.

Licences change between releases

In 2026, three labs tightened the licence on their newest large model, while Google and Meta released their newest open models under Apache 2.0.

Third-party labels can be wrong. As of 9 October 2026, LMArena’s text leaderboard lists GLM-5.3’s licence as MIT, while the file in Z.ai’s repository is the custom GLM-5.3 License, which Artificial Analysis labels correctly.¹⁶ Repositories also differ in where they keep the licence. Xiaomi’s MiMo repositories declare MIT only in the card metadata, and the Gemma 4 and Mistral Small 4 repositories have no licence file, so the card at the revision you download is your record.

Safe defaults for product teams

For a product that will self-host or redistribute weights, this order of preference keeps the legal review short.

1. Start with MIT or Apache 2.0. That set includes the top open-weight score, MiMo-V2.6-Pro at 46, and the best model under 40B parameters, Qwen3.8-27B at 34.
2. Prefer Apache 2.0 where patents matter to you. It includes an express patent licence, and MIT doesn’t mention patents.
3. Read a custom licence for its thresholds and output clauses first. Work out whether you run, or might run, a Model as a Service or AI Work Assistant business under its definition. Compare your group’s revenue and your product’s monthly users with each line, then check whether it limits models trained on the outputs.
4. Avoid Modified MIT for anything that could outgrow it. A company that passes US$20 million in monthly revenue loses its rights to Mistral Medium 3.5 and Devstral 2, including fine-tunes built on them.
5. Treat Llama 4 as a legacy choice. Its licence adds attribution, an acceptable-use policy and the EU exclusion, and Apache 2.0 models a fraction of its size score higher, such as Qwen3.8-27B at 34 and Muse Glimmer at 17.
6. Don’t plan around an unpublished licence. Mistral’s open models have shipped under both Apache 2.0 and Modified MIT, so wait for Mistral Large 4’s licence file before you commit to it.

A checklist before you ship

• Record the repository, the revision hash and the licence file you relied on, and keep a copy of each.
• Download from the organisation the model card names, and check that Hugging Face shows it as verified.
• Read the licence file itself, since tags, leaderboards and summaries like this one can be wrong or out of date.
• Measure any revenue line the way the licence does, whether that is group or company revenue, per month or per 12 months.
• Write down whether your product counts as a Model as a Service or, for Qwen, an AI Work Assistant under each definition.
• Check the acceptable-use policy against your use case, especially for regulated professions, critical infrastructure, users under 18 and companies based in the EU.
• If you ship the weights, including inside an app, include the licence text and notices, and under Apache 2.0 carry over any NOTICE file and mark the files you changed.
• If you distribute a model trained with Llama 4 or its outputs, start its name with “Llama”.
• If a third party hosts the model for you, check that the licence covers that route. Kimi K3’s licence names Moonshot’s certified inference partners.
• Repeat the check whenever you move to a new version of a model.
• Have a lawyer review any custom licence before launch.

Frequently asked questions