Open models, mid-2026: who's leading today and why the license matters more than the podium

Tracking open-weight model releases in 2026 has become a full-time job: almost every month a new model claims the top spot on some leaderboard. For anyone who has to decide on an AI architecture in a company, though, the useful question isn't "who's winning today," but what has changed structurally — and which part of that change actually matters for compliance and data control.

The podium has changed, and not by a little

In the first half of the year the top of the open-weight leaderboard shifted several times. Z.ai's GLM-5.2 (744 billion MoE parameters, 40 billion active per token, MIT license, up to a 1-million-token context) is today among the strongest on long-horizon reasoning and coding tasks. Moonshot answered with Kimi K2.7 Code, aimed at agentic coding. DeepSeek V4 Pro remains a reference point on math and code, again under a permissive license. MiniMax M3 is the first open model to combine frontier-tier software engineering capability with a 1-million-token context window and native multimodal computer-use features. Alibaba's Qwen, with the 3.5 series, remains among the strongest on multilingual performance and coding, under an Apache 2.0 license. Meta's Llama and Mistral remain solid, well-supported choices in the ecosystem, but no longer lead the leaderboards the way they did a year ago.

One fact stands out more than the scores themselves: most of these new reference models come from Chinese labs — DeepSeek, Moonshot, Zhipu/Z.ai, Alibaba — rather than from Meta or European labs. That shift is worth noting, but it shouldn't obscure the question that actually matters for a regulated company: where a model comes from is a different question from where you run it and where your data stays.

The real news isn't the podium, it's the license

Almost all the new models at the top of the leaderboards — GLM-5.2, DeepSeek V4 Pro, the Qwen family — ship under genuinely permissive licenses, MIT or Apache 2.0: free commercial use, few restrictions, no hidden thresholds tied to users or revenue. That's a concrete difference from licenses like the Llama Community License, which imposes specific conditions above certain usage thresholds.

That difference has also become a regulatory question, not just a commercial one: the European AI Office's external advisory group concluded in January 2026 that the Llama Community License does not meet the "free and open licence" requirements for the AI Act's open-source exemption. In practice, calling a model "open" is no longer enough to determine whether it falls under the regulation's lighter regime — you have to read the specific license of each release, not trust the label.

Why it matters now, not in a year

AI Act obligations for general-purpose models have applied since August 2025, but it's from August 2026 — a few weeks from now, as we write this — that penalties start applying. And it's a point many companies still underestimate: choosing an open model doesn't automatically transfer compliance responsibility to the lab that released it. The regulation explicitly places that responsibility on whoever places the system on the market or puts it into service in the European Union — in practice, on you.

What changes for a Swiss company evaluating AI today

• More choice at equal quality: with more strong model families under genuinely permissive licenses, an open-first architecture is now easier to justify on technical grounds, not just as a matter of principle.
• The license needs checking release by release: the "open" label alone no longer guarantees anything, either commercially or under the AI Act.
• Model origin and where you run it are two separate questions: an open model developed elsewhere, run on your own infrastructure or a Swiss sovereign cloud, with your data never leaving that perimeter, is a fully different choice from relying on a foreign API — even when the model comes from the same lab.
• Regulatory deadlines are getting closer: with AI Act penalties on the way, checking your architecture's compliance now costs far less than discovering a problem later.

This is exactly the work we do every week with our clients: tracking these releases, reading the licenses before they become a problem, and building an architecture where the right model runs where it needs to run — with your data staying yours, regardless of who trained the model.