Open-Weight AI License Tracker: What You Can Actually Do With Each Model

Open Weights No Longer Means One License

Downloadable weights tell you that you can run a model on your own hardware. They don’t tell you what you may do with it. That used to be a footnote, because most popular open models shipped under MIT or Apache 2.0. In 2026 the licenses split, and the split runs right through some model families: GLM-5.3-Flash is MIT while GLM-5.3 is not, and Qwen3.8-27B is Apache 2.0 while Qwen3.8-Max is not.

The models a business is likely to consider now fall into three groups.

1. Permissive — MIT or Apache 2.0. Commercial use, modification and redistribution are allowed. You keep the license notice when you redistribute the weights, and that is about it.
2. Custom — A license written by the developer. Commercial use is generally allowed, but with some mix of attribution, revenue or user thresholds, model-as-a-service clauses and a usage policy.
3. Not yet announced — The model is out on an API and weights are promised, but the license hasn’t been published. MiniMax M3.1 Flash and Mistral Large 4 sit here.

This is not legal advice. The table below summarizes vendor announcements, press coverage and third-party license guides. Where a term is “reported”, we have not seen it confirmed in the license text. Before you deploy, read the license file in the model’s own repository on Hugging Face or GitHub, for the exact version you download, and ask counsel about anything that touches your business model.

The Master Table: 15 Open-Weight Models and Their Licenses

“Your own product” means a model built into software or a service you sell, such as a feature in your app. “Model-as-a-service” means selling access to the model itself, for example as a hosted API. Source numbers refer to the References at the end of this page.

Two rows deserve a closer look. MiniMax M3 is the one custom license in the table whose terms reach a small company directly: according to a published guide to the license, an organization below $20M in yearly revenue owes MiniMax a one-time notice, and one above it needs prior written authorization. And Muse Spark Open’s attribution and usage terms apply at any size, which the next two sections cover.

Who the Custom Conditions Actually Hit

Read the custom licenses side by side and most of the friction lands on a narrow group: companies that resell access to the model itself at scale. A 40-person firm running a model for internal document work, or inside a customer-facing app, sits well below most of these lines. Meta’s license says so explicitly for Muse Spark Open, and Alibaba’s exempts internal use of Qwen3.8-Max.

The attribution duty is easy to underestimate. If a software firm builds a client portal on Muse Spark Open, the portal needs a visible “Built with Muse Spark Open” somewhere in the interface, documentation, website or settings. The duty follows the model into fine-tunes and, for Muse, into models trained on its outputs. MiniMax M3’s “Built with MiniMax M3” requirement also covers fine-tuned derivatives. Neither is a burden, but both are easy to forget when a prototype becomes a product.

Muse Spark Open and Regulated Professions

Meta released Muse Spark Open’s weights on October 15, 2026 and calls it “the most capable US open-weight model to date”. Its license is generous in most respects: free use, modification, fine-tuning, distillation and redistribution, commercial use internally and in your products, no claim on outputs, and none of the user caps or geographic limits of older Llama licenses.

The usage policy is where clinics, accounting firms and law practices should slow down. According to coverage of the license, it bans uses with a risk of death or bodily harm, including operating critical infrastructure, and the unauthorized practice of regulated professions, naming finance, law and medicine. It also bans military, warfare, nuclear and espionage uses, and the usual illegal and harmful ones.

The operative word in the professions clause appears to be “unauthorized”. A family-medicine clinic summarizing charts that a physician reviews, or a law firm drafting documents a lawyer signs, looks different from a public chatbot that gives medical or legal advice with no licensed professional behind it. That reading is ours, not Meta’s, and it is not legal advice. If your work sits anywhere near the line, have counsel read the policy against your actual workflow. Utilities and companies whose software controls physical equipment should read the bodily-harm clause the same way.

MIT and Apache 2.0 contain no usage policy at all. That doesn’t relax your professional obligations, which apply whatever model you use, but it removes a second set of rules written by the model vendor.

The Simplest Models to Deploy: MIT and Apache 2.0

If you want no license questions beyond keeping a notice file, start with the permissive group. The five below are also the ones whose sizes we can map to Mac Studio memory. Footprints are rough figures at 4-bit quantization; the suggested uses are Faraday’s editorial view. Where a 1- to 3-bit build changes the fit, the Fits column says so; those lower-bit builds trade some quality, most noticeably on long or precise work.

Three more Apache 2.0 options round out the group: Google’s Gemma 4 family, up to a 31B dense model; Cohere’s Command A+, 218B total with 25B active; and Reflection AI’s Beam 1.0, 501B total with 23B active, whose Apache 2.0 license is vendor-stated and whose weights were only reported on October 16. For Beam, check that the license file in the repository matches the announcement before you build on it.

The License Trend: Tighter at the Top

The pattern across 2026 is consistent. Developers keep small and mid-sized models permissive and put conditions on the flagship. A few have moved the other way.

Two practical lessons follow. First, a license check belongs to a version, not a family. A team that approved GLM-5.2 under MIT in June cannot assume GLM-5.3 carries the same terms. Second, the custom terms so far are aimed at large companies reselling model access, not at businesses running a model for their own work. That could change with any release, which is why MiniMax M3.1 Flash, due as open weights on October 23, stays in the “not yet announced” group until its license file appears.

Before You Deploy: A License Checklist

Work through these before any open-weight model goes live.

1. Read the license file itself — Open the LICENSE file in the model’s Hugging Face or GitHub repository. Blog posts, press coverage and this tracker are summaries.
2. Match the exact version — Check the license for the precise model and size you download. GLM-5.3 and GLM-5.3-Flash, or Qwen3.8-27B and Qwen3.8-Max, carry different terms.
3. Place the model in a group — Permissive, custom or not yet announced. If there is no license file yet, don’t deploy it in production.
4. Describe your use in one line — Internal tool, feature inside a product you sell, or access to the model sold as a service. Most custom clauses turn on this distinction.
5. Compare your numbers to the thresholds — Note how each license measures revenue (yearly, monthly, trailing, any 12 months) and whether it counts your company or only a model-as-a-service business.
6. Send any required notices — MiniMax M3 asks for a one-time notice, or prior written authorization above $20M in yearly revenue. Keep a copy of what you sent.
7. Plan the attribution — Decide where “Built with Muse Spark Open” or “Built with MiniMax M3” will appear before the product ships, and carry it into fine-tuned versions.
8. Test the usage policy against your industry — Health, legal, financial and infrastructure work should get a counsel review against Muse’s policy, or a permissive model instead.
9. Keep the data question separate — A license governs the weights, not where your prompts go. Running the model on your own hardware is what keeps client data in the building; see how cloud AI data can train competitors’ models.
10. File the license with the weights — Save the license text, version and download date alongside the model, so you can show which terms applied when you deployed.

Where Faraday Fits

Faraday Machines installs open-weight models on Mac Studio hardware in Toronto. Four of the five models on our standard list, DeepSeek V4.1 Flash, GLM-5.3-Flash, Qwen3.8-27B and MiMo-V2.6-Flash, are MIT or Apache 2.0. The fifth, MiniMax M3, uses the minimax-community license, so we flag its attribution and notice terms to clients who choose it. Kimi K3, GLM-5.3 and MiMo-V2.6-Pro are available on larger clusters; the first two carry custom licenses.

Muse Spark Open is a different hardware class. The supported Mac configurations reported at launch are two M5 Ultra Mac Studios with 512 GB each, or four with 256 GB [4]. For a broader comparison of what these models do, see Open Source AI Models Compared.

References