Two models can both call themselves "open" and hand you completely different rights. One lets you fine-tune it, wrap it in a product, and sell that product with no restriction beyond keeping a copyright notice in place. The other's free grant quietly expires the moment your product crosses a user count nobody mentioned in the launch tweet — and by then, renegotiating is a business problem, not a code change. The fix costs about five minutes: read the actual license file, not the announcement, and check it against four specific things, every time.
What "open-weight" doesn't promise
"Open-weight" only means the trained parameters are downloadable — it says nothing about the terms attached to them. A genuine open-source license (Apache 2.0 or MIT, the same terms most of the software industry already trusts) grants blanket rights to use, modify, redistribute, and sell. A "community license" or "acceptable use" license can sit behind the identical download link with a completely different set of conditions: a user-count ceiling, a revenue threshold, a ban on using the model's own outputs to train a rival, or a naming requirement once your product gets big enough to notice. The download step looks identical either way — the difference lives entirely in a document most people never open.
Check the license before you ship
- Open the model's card on Hugging Face or its GitHub repo and read the LICENSE file directly. A launch blog post is marketing copy, and it can describe terms that change before the model does.
- A real Apache 2.0 or MIT grant — both OSI-approved, both give blanket rights to use, modify, redistribute, and sell — versus a custom "community license" or "research license" the vendor wrote itself.
- Search the text for "monthly active users" or "revenue." Some "open" licenses expire your free grant entirely past a user count; others only add an attribution requirement.
- Some licenses ban using the model's own outputs to train or improve a different model — a real constraint if your pipeline does any distillation or synthetic-data generation.
- The same lab can ship a restrictive preview license and a permissive full release ten weeks apart, or the reverse. A license you checked in a prior release cycle is not evidence for this one.
Four real 2026 licenses, side by side
The differences aren't hypothetical. Four models released this year, all marketed as "open," land in four different places once you actually read the document. DeepSeek's V4 Pro ships under a true MIT license — commercial use, redistribution, and training other models on its outputs are all unrestricted beyond keeping the copyright notice intact. Meta's Llama 3 Community License permits commercial use, but its free grant expires outright once a licensee's products cross 700 million monthly active users in the preceding month, and it separately bars using Llama's outputs to improve any other large language model. Moonshot's Kimi K2 ships under a Modified MIT License that stays permissive at any size — the only added condition is displaying "Kimi K2" in your product's interface once you clear 100 million monthly active users or $20 million in monthly revenue. And Tencent's Hunyuan Hy3 changed its own answer mid-year: its April preview shipped under a restrictive custom license that excluded the EU, UK, and South Korea outright, and its July full release dropped that entirely for a plain Apache 2.0 grant.
What four 2026 "open" releases actually let you do
| DeepSeek V4 Pro MIT | Kimi K2 Modified MIT | Llama 3 Community License | Hunyuan Hy3 Apache 2.0, current release | |
|---|---|---|---|---|
| License family | MIT | Modified MIT | Llama 3 Community License | Apache 2.0 |
| Commercial use | Unrestricted | Unrestricted | Allowed, with a cap | Unrestricted |
| Where the grant changes | Never | Attribution required past 100M users or $20M monthly revenue | Free grant expires past 700M monthly users; a separate license is required beyond it | Was geo-restricted at its earlier preview; dropped entirely at the current release |
| Training other models on its outputs | No restriction | No restriction | Explicitly banned for other large language models | No restriction |
Where this reading goes wrong
Five ways this check gets skipped
None of this is abstract for this beat specifically: a license's field-of-use clause is the same territory as the distillation dispute this newsroom has covered directly — Washington's unproven accusation that Moonshot distilled Anthropic's models to build Kimi K3 turns on exactly the question a license's training-restriction clause is meant to settle in advance: whose outputs can legally train whose next model. (A license with no field-of-use restriction doesn't make an unauthorized-access claim like that one go away — it only settles what you're allowed to do with a model you already have lawful access to.) If the next question is whether to self-host the model at all once its license clears, our guide to deciding between self-hosting and an API picks up from here.
- "Open-weight" doesn't mean open-source — read the actual license file, never the launch post.
- DeepSeek's V4 Pro ships true MIT; Meta's Llama 3 grant expires past 700 million monthly users.
- Moonshot's Kimi K2 only requires attribution past 100 million users or $20 million monthly revenue.
- Tencent's Hunyuan Hy3 went from a restricted preview license to Apache 2.0 in ten weeks.
- Caveat: a license you checked on a past release says nothing about the version you're deploying now.