Apache 2.0, MIT or 'community': what open model licenses let you ship

In the past week, three large open models landed with three different licenses. "Open" on a model page can mean "do anything", or it can mean "free, but with conditions". This guide walks through the licenses you'll actually meet and what each lets you ship commercially, based on the license texts.

Not legal advice. This is a plain-English reading of public license texts as of Oct 22, 2026. If a product, a client contract or real money depends on it, read the license file yourself and ask a lawyer.

The past week's releases and their licenses

• Beam. Reflection says it released Beam's weights on Hugging Face "under an Apache 2.0 license" in its launch post dated Oct 16, 2026. The model card lists 501.3B total parameters and 23.4B active, and the repo includes the full Apache 2.0 LICENSE file. The card also says it "does not modify or restrict the model's license".
• Ornith-1.5-122B. Its model card declares license: mit. Its GGUF, FP8 and NVFP4 versions went up on Oct 20, 2026, per Hugging Face. One catch: the card links to a LICENSE file that wasn't in the repo when we checked on Oct 22. MIT's one condition is keeping the notice, so check for that file before you redistribute. We compared Ornith with Flint in Flint vs Ornith local coding models.
• MiniMax M3.1 Flash. The weights went public on Hugging Face on the evening of Oct 22, 2026, under the MiniMax Community License. We cover the model in MiniMax M3.1 Flash. The license terms are below.

MIT: do almost anything, keep the notice

The MIT License grants the right "to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies" without restriction.

There's one condition: the copyright and permission notice must be "included in all copies or substantial portions". The software comes "as is", with no warranty.

What it means for you: you can self-host, fine-tune and sell a product built on an MIT model. If you redistribute the weights (including a quantized copy), ship the license with them. DeepSeek's V4.1 Flash weights are MIT, per its card. See our DeepSeek V4.1 Flash piece.

Apache 2.0: permissive, plus a patent license

Apache 2.0 also allows use, modification, sublicensing and distribution. Section 3 adds a patent license, which MIT doesn't spell out.

The conditions when you redistribute (Section 4):

• Give recipients a copy of the license.
• Mark modified files with "prominent notices" that you changed them.
• Keep the copyright, patent, trademark and attribution notices.
• If there's a NOTICE file, pass its attribution notices along.

Two more lines matter:

• Patent litigation: if you sue claiming the work infringes a patent, your patent license from that contributor ends (Section 3).
• Trademarks: the license doesn't let you use the licensor's names or marks, beyond describing where the work came from (Section 6).

What it means for you: commercial use is fine. Fine-tunes and quantized releases need the license file and a note that you changed things.

OpenMDW-1.1: a permissive license written for weights

The OpenMDW-1.1 license was written for "Model Materials": the model's architecture and parameters, plus related data, docs and code.

It grants permission "to deal in the Model Materials without restriction", including under "copyright, patent, database, and trade secret rights".

• Condition: if you distribute any part, keep a copy of the agreement and the copyright and origin notices.
• Outputs: it "does not impose any restrictions or obligations" on outputs.
• Lawsuits: if you sue claiming the materials infringe a patent or copyright, your rights end (unless you were sued first).
• Your job: you're "solely responsible" for clearing other people's rights.

NVIDIA's Nemotron 3.5 Super uses OpenMDW-1.1. See Nemotron 3.5 explained.

Llama-style licenses: free, with a credit line and a size cap

Meta's Llama 4 Community License (effective Apr 5, 2025) is the template many people mean by "Llama-style".

It grants a royalty-free right to use, copy, modify and distribute. Then it adds conditions:

• Credit: if you distribute the model or a product that contains it, include the agreement and "prominently display 'Built with Llama'".
• Naming: if you use Llama or its outputs to train a model you distribute, its name must start with "Llama".
• Notice file: keep the Llama 4 attribution notice in a "Notice" text file.
• Size cap: if your products had more than 700 million monthly active users in the month before the Llama 4 release date, you must request a license, which Meta may grant "in its sole discretion".
• Use policy: you must follow the Acceptable Use Policy. Among other things, it bans military and warfare uses.
• EU and multimodal: the use policy says rights to Llama 4's multimodal models aren't granted to people or companies based in the EU. End users of a product built with them aren't affected.
• Governing law: California.

What it means for you: for most indie builders, commercial use is allowed. Budget for the "Built with Llama" credit, the naming rule for fine-tunes, and the use policy.

"Community" licenses like MiniMax's: free for some uses, conditions for business

The MiniMax Community License for M3.1 Flash starts from MIT-like wording but limits the grant to non-commercial purposes. Commercial use comes with conditions:

• Credit: "prominently display 'Built with MiniMax M3.1-Flash'" on a related website, interface, blog post, about page or docs.
• Over $20 million a year in revenue: you need "separate, prior written authorization" from MiniMax.
• Below that: a one-time notice to MiniMax, as the license describes.
• "Commercial Use" is broad. It includes paid products that rely on the model, commercial use of APIs serving it (cloud or hosted), and commercial deployment of fine-tuned or modified versions.
• Banned uses: an appendix rules out illegal content, any military purpose, harming minors, harmful false information, and discrimination or hate.

What it means for you: experimenting is free. Before you charge money for anything that runs the weights, add the credit line and send the notice. Calling MiniMax's own API is a different deal: the API terms apply, not this license.

Quick comparison

A 5-minute license check before you ship

1. Open the LICENSE file in the exact repo you download. A quantized repo by someone else might show a different tag than the original.
2. Check the original model's license too. The licenses above follow copies and derivatives. A GGUF re-upload doesn't drop the original terms.
3. Look for extra documents. Use policies (Llama) or prohibited-use appendices (MiniMax) can be part of the deal.
4. Note the thresholds: user counts, revenue, or region.
5. Add credits on day one. An about page or docs line is cheap. Retrofitting it later is easy to forget.

If you're planning to self-host any of these, start with the hardware math in how much VRAM you need, and if it doesn't fit, see running a model bigger than your RAM.