AI Labs Outside the US and Their Open-Weights Licences

Several large AI labs outside the United States publish models as open weights that anyone can download. DeepSeek, Alibaba’s Qwen team, Moonshot AI, Z.ai, MiniMax and Mistral publish their newest large models this way; ByteDance publishes some models but keeps its Seedance video model closed. What “open” lets you do differs from one licence to the next, and that is the part worth reading before you build on any of them.

Why open weights matter here

This site separates frontier labs whose models you rent from an open-weights layer whose models you can download, run and keep, trailing the frontier by months, not years. That layer is why AI is not simply owned by a few companies; see who builds AI and open or closed. Open weights are not open source: you can run and adapt the model, but not rebuild it.

A September 2026 example shows the difference, and how little warning an API gives. On 10 September DeepSeek announced that “Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates.”DeepSeek, DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient, 10 Sep 2026, read at source 22 Sep 2026: “Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates.” The news post still carries this sentence. Before that date arrived, DeepSeek reversed the plan. Its change log says: “In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged.” Its pricing page still lists deepseek-v4-pro, served by DeepSeek-V4-Pro-0813, at prices of its own.DeepSeek, Change Log, entry dated 2026-09-10, and Models & Pricing, both read at source 22 Sep 2026. Internet Archive copies of the change log lack the reversal on 11 Sep 2026 and carry it on 13 Sep 2026. So anyone calling that API name had days’ notice of a model swap they had not asked for, and then had to catch a second notice saying it was off. Anyone who downloaded the V4-Pro weights did not need to read either one.Reasoning, September 2026 — follows from the two notices above and from the V4-Pro weights remaining on DeepSeek’s Hugging Face page, read at source 22 Sep 2026.First-hand: until 22 Sep 2026 this paragraph said “Anyone calling that API name now gets a different model.” That was wrong. DeepSeek had withdrawn the switch in its change log before it was due; this page took the plan from the 10 Sep news post, which still carries the original sentence and not the reversal.

The labs, one at a time

DeepSeek (China). Its privacy policy names the operator as “Hangzhou DeepSeek Artificial Intelligence Co., Ltd., with its registered address in China”.DeepSeek, Privacy Policy, last updated 10 Feb 2026, read at source 18 Sep 2026. Its newest release, DeepSeek-V4.1-Flash, is dated 10 September 2026, and its model card’s licence section reads: “This repository and the model weights are licensed under the” MIT License.DeepSeek, API news: “DeepSeek-V4.1-Flash Release 2026/09/10”; DeepSeek-V4.1-Flash model card. Both read at source 18 Sep 2026.

Qwen (Alibaba). Alibaba released the weights of its flagship on Hugging Face and on ModelScope, which it describes as “an Alibaba-initiated open-source community”. Its release log dates Qwen3.8-2.4T-A95B to 12 August 2026 and the smaller Qwen3.8-27B to 14 August, and says: “For the first time, Qwen3.8 brings a Qwen-Max-class model to open release.”Alibaba Cloud, Alibaba Unveils Qwen3.8-27B and Releases Weights of Qwen3.8 Flagship Model; Qwen, Qwen3.8 on GitHub: “2026-08-12: Qwen3.8-2.4T-A95B is now available on” and “2026-08-14: Qwen3.8-27B is now available on”. Both read at source 18 Sep 2026. The two carry different licences, and a later model, Qwen3.8-Flash-Next, carries a third; see the next section.

Moonshot AI (Kimi). Its About page gives a Beijing address and says the company was founded in spring 2023. It lists Kimi K3 on 16 July 2026, “a 2.8T-parameter model” released “under the Kimi K3 License”.Moonshot AI, About: “Address 76 Zhichun Road, Haidian District, Beijing JD Technology Building 13F”; home page: “2026-07-16 Kimi K3”; Kimi-K3 model card. All read at source 18 Sep 2026.

Z.ai (GLM models). Its GitHub page links each GLM download on ModelScope under the organisation name ZhipuAI, and its privacy policy says the international service is run from Singapore. It publishes GLM-5.3 and a smaller GLM-5.3-Flash, “With 320B total parameters and just 18B active parameters”.Z.ai, GLM-5 on GitHub, linking “https://modelscope.cn/models/ZhipuAI/GLM-5.3”; GLM-5.3-Flash model card; Privacy Policy: “We generally provide the Services from Singapore”. All read at source 18 Sep 2026.

MiniMax. It calls itself “a global AI foundation model company. Founded in early 2022”. It publishes MiniMax-M3, a multimodal model, and MiniMax H3, a video model that “can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds”. The H3 licence gives its release date as 2 August 2026 and is “governed by the laws of the Hong Kong Special Administrative Region”.MiniMax, About; MiniMax-H3 model card; MiniMax H3 Community License Agreement: “MiniMax H3 release date/License date: August 2, 2026.” All read at source 18 Sep 2026.

ByteDance Seed. Seedance 2.5, its video model, is “built for 30-second storytelling with precise reference control and powerful editing capabilities”; its page offers “Get API” and “Try now”, not a download. The ByteDance-Seed organisation on Hugging Face listed 59 models on 18 September 2026, none of them Seedance. It does publish Seed-OSS-36B, a language model, under Apache 2.0.ByteDance Seed, Seedance 2.5, read in a browser 18 Sep 2026; Hugging Face, ByteDance-Seed model listing, read at source 18 Sep 2026; Seed-OSS-36B-Instruct model card: “We release this series of models to the open-source community under the Apache-2.0 license.” Prompting it is covered on the Seedance cheatsheet.

Mistral AI (France). Its privacy policy says “Mistral AI is a French company incorporated in Paris”. Mistral Medium 3.5, announced on 22 May 2026, was “Released as open weights, under a modified MIT license.” Mistral Small 4 is under Apache 2.0.Mistral AI, Privacy Policy; Remote agents in Vibe. Powered by Mistral Medium 3.5, dated May 22, 2026; Mistral Small 4 model card: “Open-source license for both commercial and non-commercial use”. All read at source 18 Sep 2026.

Licence names from each model’s own card and LICENSE file on Hugging Face, and the per-organisation listings on Hugging Face, all read at source 18 Sep 2026.

“Open weights” is not one licence

MIT and Apache 2.0 are standard permissive licences that allow commercial use, with conditions such as keeping the licence notice. The other labs write their own terms, and those differ in ways that matter to a business.Mistral AI, Mistral Small 4 model card, read at source 18 Sep 2026, describing Apache 2.0 as “Open-source license for both commercial and non-commercial use”.

Conditions once you are large. The Qwen3.8-Max licence, covering the 2.4-trillion-parameter model, wants the model name shown in products with “more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue”. A model API or work-assistant business whose “aggregate revenue of the licensee and its affiliates exceeds US$50,000,000” in twelve months must “obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose”. The Kimi K3 licence sets its API-business threshold at 20 million US dollars.Qwen, Qwen3.8-Max License; Moonshot AI, Kimi K3 License: “the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars”. Both read at source 18 Sep 2026.First-hand: until 22 Sep 2026 this paragraph quoted the display clause only as far as “more than 100,000,000 monthly active users”, leaving out its monthly-revenue alternative. Full clause re-read in the Qwen3.8-Max License at source 22 Sep 2026. Alibaba’s announcement stresses Apache 2.0, which is true of the 27B model, not the larger one; read the file next to the weights, not the headline.Reasoning, September 2026 — compares the Alibaba Cloud post (“Released under the Apache 2.0 license”, said of Qwen3.8-27B) with the two model cards, all read at source 18 Sep 2026.

A separate deal for any API business. Qwen3.8-Flash-Next, on Hugging Face since 24 August 2026, carries a third Qwen licence, the Qwen Community License 1.0. Its display clause is the same as the Qwen3.8-Max one. Its second clause drops the revenue threshold: if the licensee “conducts a Model as a Service or AI Work Assistant business, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.” Internal use that makes nothing available to third parties is exempt.Qwen, Qwen Community License 1.0 for Qwen3.8-Flash-Next, read at source 22 Sep 2026; date from the Hugging Face API listing for Qwen, read 22 Sep 2026: “2026-08-24T08:24:59.000Z Qwen/Qwen3.8-Flash-Next”.

A security review. The GLM-5.3 licence adds one condition beyond keeping the notice: a company running a model API business with more than 10 billion US dollars of revenue over twelve months must pass Z.ai’s security review before commercial use. “The scope and method of the security review shall be reasonably determined by Z.AI.” GLM-5.3-Flash, by contrast, is MIT.Z.ai, GLM-5.3 License: “exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months”; GLM-5.3-Flash model card: “license: mit”. Both read at source 18 Sep 2026.

A revenue cap. Mistral Medium 3.5’s modified MIT licence gives no rights at all to larger companies: “You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month.”Mistral AI, Modified MIT License, read at source 18 Sep 2026: “exceeds $20 million (or its equivalent in another currency) for the preceding month”.

Non-commercial by default, and limited by territory. The MiniMax-M3 licence grants rights “to deal in the Software for non-commercial purposes”; commercial use requires crediting MiniMax and a notice to it, or its written authorisation above 20 million US dollars of yearly revenue. The H3 video licence goes further: its “Excluded Territories” mean “the European Union, the United Kingdom, the Republic of Korea and the United States of America”.MiniMax, MiniMax Community License: “if such products and services generate more than 20 million US dollars (or equivalent in other currencies) in yearly revenue”; MiniMax H3 Community License Agreement. Both read at source 18 Sep 2026.

Research only. Update, 22 September 2026: Qwen’s image model Qwen-Image-2.1 is under none of the three Qwen licences above. Its Qwen Research License Agreement, marked “Release Date: September 20, 2026”, grants rights “FOR NON-COMMERCIAL PURPOSES ONLY” and defines “Non-Commercial” as “for research or evaluation purposes only”. Commercial use needs “a separate commercial license” from the licensor, Hangzhou Tongyi Laboratory Technology Co., Ltd. The same lab now publishes under Apache 2.0 and under research-only terms.Qwen, Qwen Research License Agreement and Qwen-Image-2.1 model card, both read at source 22 Sep 2026.

Where to find them safely

Each lab above publishes its weights under its own Hugging Face organisation, listed in the table; Qwen, Z.ai and MiniMax also link copies on ModelScope. Start from the lab’s page and follow its link. A Hugging Face search for “seedance” on 18 September 2026 returned 17 model repositories, none owned by ByteDance-Seed.Hugging Face, model search for “seedance”, read through the Hugging Face API 18 Sep 2026. A famous name on a repository does not make it the lab’s release. For running a model yourself, see local or cloud.

What is genuinely uncertain

The hosted apps and APIs are separate from the weights, and their policies differ. DeepSeek’s says: “The Personal Data we collect from you may be stored on a server located outside of the country where you live.” It names the People’s Republic of China as where it stores that data, and lists “the right to opt-out of using your Personal Data for training our models or optimizing our technologies.”DeepSeek, Privacy Policy, last updated 10 Feb 2026, read at source 18 Sep 2026. Z.ai’s international service says data “is generally processed in Singapore”. Kimi’s names NOVASCENT PRIVATE LIMITED as “the provider and controller” and says user content is processed to improve the services, “including training and optimizing our artificial intelligence” models.Z.ai, Privacy Policy; Kimi, Privacy Policy, last updated 13 Aug 2026. Both read at source 18 Sep 2026. These policies cover each lab’s own service, not weights you run on your own hardware.Reasoning, September 2026 — a privacy policy covers the lab’s own service; see what happens to what you type.

The one territorial restriction found in these primary sources is MiniMax’s, and the licence gives no reason for it. It says only: “We will continuously evaluate the applicable laws, regulations and compliance requirements for the Excluded Territories.”MiniMax, MiniMax H3 Community License Agreement, read at source 18 Sep 2026. This page does not guess at the reason.

What this page could not verify

• Any benchmark claim. Each card compares itself with rivals; none of those scores was checked here, and this page ranks nothing.
• Training data. This page did not assess what any of these models was trained on.
• Where MiniMax and ByteDance Seed are based. The pages read here give no headquarters address for either.
• Release dates for MiniMax-M3 and Mistral Small 4, and the day the GLM weights went up. Z.ai’s developer release notes list “2026-08-18 GLM-5.3” and “2026-08-26 GLM-5.3-Flash”; those are API release dates, and need not be the day the weights were published. For the other two, no date was found on a page that could be read; the table gives none.Z.ai, New Released release notes, read at source 22 Sep 2026.
• Why MiniMax excludes four territories. The licence does not say.
• Whether this list is complete. Other labs outside the US publish weights; this page covers the seven named in its brief that could be checked at source.