MIT-Licensed Models: What You Can Actually Build On

Licence is the only column in a model comparison that can invalidate your product retroactively. A benchmark score being wrong costs you some quality. A licence term you did not read costs you a renegotiation at exactly the point your business becomes worth negotiating over.

In 2026 the open-weight field splits cleanly on this. GLM-5.2 and both DeepSeek V4 variants are MIT. The Kimi models are not. That distinction decides more architecture than any leaderboard does.

What MIT actually gives you

MIT is about as permissive as a licence gets. Use commercially, modify, redistribute, sublicense, build a product on top — the only obligation is preserving the copyright notice and the licence text.

Critically, there is no revenue threshold, no monthly-active-user ceiling, no field-of-use restriction and no requirement to negotiate a separate agreement once you reach a certain size. What you can do on day one is what you can do at a hundred million dollars of revenue.

That absence of a cliff is the whole value. Restrictive terms in model licences are almost always structured to activate at scale, which means they are invisible during evaluation and expensive at success. MIT removes the category. Open weights versus open source covers why the two phrases are not synonyms, and why labs choose their words carefully.

The models on the MIT side

GLM-5.2, released by Z.ai on 13 June 2026, is the capability leader of this group. It is a mixture-of-experts model of roughly 744B total parameters with about 40B active, a 1M-token context window, a 128K maximum output length, and two reasoning effort levels. It reports 62.1 percent on SWE-bench Pro and 81.0 on Terminal-Bench 2.1, and sits around 51 on the Artificial Analysis Intelligence Index. List pricing is around $1.40 per million input tokens and $4.40 output.

DeepSeek V4 Pro, released 23 April 2026, is 1.6T total with about 49B active, 1M context, around 44 on the same index, and reports 80.6 percent on SWE-bench Verified. It lists at roughly $0.44 in and $0.87 out — noticeably cheaper than GLM-5.2 for a model in the same tier.

DeepSeek V4 Flash comes from the same 23 April release, activates about 13B per token, keeps the 1M context, and lists around $0.14 in and $0.28 out. A 0731 refresh landed on 30 July 2026. It is the cheap end of the MIT field. The Pro versus Flash comparison covers when the smaller one is sufficient.

Qwen sits alongside these under permissive terms on most tiers, and its dense 27B — released 20 April 2026, single GPU, 77.2 percent on SWE-bench Verified — is the practical choice when you need something you can genuinely own. Check the specific tier's terms rather than assuming the family is uniform.

The model that is not MIT, and why it matters

Kimi K3 is the strongest open-weight model of the group by most measures: 2.8T total with roughly 104B active, 1M context, around 57 on the Artificial Analysis index, first on Arena Frontend Code at 1,679, 88.3 on Terminal-Bench 2.1 and 81.2 on FrontierSWE. Moonshot published roughly 1.6TB of weights on 27 July 2026, eleven days after the 16 July release.

But it ships under Moonshot's own terms rather than MIT, including a revenue threshold that triggers a separate agreement for model-as-a-service businesses. Kimi K2.6, released 20 April 2026, is likewise not MIT.

For an internal engineering team, none of that bites — you are not reselling inference and no threshold applies. For a company whose product is inference, it is a gate that appears at exactly the point you succeed, and it is far cheaper to resolve before you build than after. The Kimi K3 guide covers the model itself; the open-weight comparison covers the wider licence spectrum.

What a licence does not give you

MIT tells you what you may legally do. It says nothing about what you can practically do, and conflating the two is the most common mistake in this area.

GLM-5.2 at roughly 744B total parameters must be held entirely in memory regardless of how few experts a given token activates. DeepSeek V4 Pro at 1.6T is worse. Both are multi-accelerator deployments. "MIT-licensed" and "you can run it yourself" are separate claims, and only the first is guaranteed by the licence. Dense versus mixture-of-experts covers why sparse models are so much easier to rent than to own.

MIT also gives no warranty, no support commitment and no guarantee that a future release stays permissive. A licence applies to the artefact it shipped with, and the next version is a fresh decision by the lab.

The mitigation is boring and effective: archive the weights and the licence text of any version you depend on. If a future release changes terms, the version you already have keeps the terms it shipped under.

When the licence genuinely decides it

Four situations, and outside them the licence column is mostly noise.

If you resell inference — an API product, a hosted agent, anything where customers pay you for model output — the model-as-a-service clauses in non-MIT licences apply directly to you, and MIT removes a legal conversation entirely.

If you ship weights inside a product that customers install, redistribution rights are the binding term and permissive licensing is close to mandatory.

If you fine-tune and want to publish the result, you need the right to distribute derivatives. MIT grants it; custom licences vary and often attach conditions.

If you operate in a regulated environment where legal must sign off on every third-party dependency, MIT is a licence your counsel has approved a thousand times before. That alone can save a quarter of procurement time, which is a real cost even though it appears on nobody's benchmark.

How to use this in practice

Settle the licence question before you evaluate capability, not after. It is a filter, not a tiebreak, and running a three-week evaluation on a model your business model cannot use is pure waste.

Write down which of the four situations above applies to you. If none do, drop the licence column and choose on capability, price and context — and take Kimi K3 seriously, because it is the strongest of the group and its terms will never affect you.

If any of them do apply, your shortlist is GLM-5.2 when you need capability, DeepSeek V4 Pro when you want most of it for less, and DeepSeek V4 Flash when the task is bounded and volume is high. The provider selection checklist covers the rest of the due diligence.