--- title: "Open weight or open-washing? What Qwen3.8-Max, DeepSeek V4 and Europe's EUROPA model teach us" description: "Three stories from the same week in August 2026 show why "open" is no longer a guarantee: Qwen3.8-Max announced without weights or a license, DeepSeek V4 already open under MIT, and the European EUROPA consortium built to be compliant from the contract stage." date: 2026-08-10 tags: [Open models, News, Licenses] url: "https://ai.malagoli.me/en/blog/open-weight-o-open-washing-agosto-2026" locale: en_US image: "/blog/open-weight-o-open-washing-agosto-2026.svg" --- # Open weight or open-washing? What Qwen3.8-Max, DeepSeek V4 and Europe's EUROPA model teach us In the same week, Alibaba announces an "open" model with no weights and no license, DeepSeek publishes its own under MIT, and the European Union picks who will build its sovereign open model. Three very different degrees of openness. *** Three stories from the same week in early August 2026 show three very different degrees of what "open" actually means for AI models. Alibaba presented Qwen3.8-Max as its flagship model's return to open source — but with no downloadable weights and no published license on the day of the announcement. DeepSeek, over the same stretch, had already published V4-Flash's weights under the MIT license, the most permissive form there is. And the European Commission selected the consortium that will build EUROPA, an open model designed to be AI Act-compliant from the first day of training. For anyone evaluating an open model for their company, the difference between these three cases isn't a technical detail — it's the question to ask before choosing. ## Qwen3.8-Max: announced as open, but the weights aren't there yet On August 3, 2026, Alibaba presented Qwen3.8-Max, a 2.4-trillion-parameter model positioned as the Qwen-Max line's return to open source, after its most recent flagship models had stayed closed. The problem is that the announcement came with no actual release: no model page on Hugging Face, no downloadable weights, no published license — only a promise that weights would follow "within about a week". The developer community's reaction was blunt: The New Stack called it "an API business model wearing an open source jacket for the launch photo". ## The problem isn't just the delay — it's what the license will say once it lands The most delicate point isn't when the weights will arrive, but what they'll actually let you do. Some developers, reading drafts circulating online, flagged the risk that the license might explicitly exclude use — and even download — of the model from the United States, the European Union, the United Kingdom and South Korea, a reading Alibaba has neither confirmed nor denied. This isn't a theoretical concern: another flagship Chinese model, MiniMax H3, has already shipped "open" weights whose license explicitly excludes local deployment in exactly those four jurisdictions. For a Swiss company evaluating self-hosting an "open" model, this is precisely the difference between "the weights can be downloaded" and "you can legally run them here" — a check to make before building a project on top of it, not after. ## DeepSeek V4: the other extreme — weights already published under a permissive license Over the same stretch of weeks, DeepSeek published the weights of DeepSeek-V4-Flash-0731 on July 31, 2026 under the MIT license — the most permissive open source license in circulation, with no geographic restrictions and no commercial-use constraints. The model, with its 284 billion parameters and a 1-million-token context window, is downloadable and self-hostable right away. The more capable variant, the 1.6-trillion-parameter V4-Pro, remains in preview for now. It's the useful reminder in this comparison: a genuinely open model isn't announced, it's published — weights, license and technical documentation together, the same day. ## Europe is trying to build an open model compliant from the contract onward In the same week, the European Commission selected the EUROPA consortium, led by Italian company Domyn together with the German research organization Fraunhofer-Gesellschaft, as the winner of the Frontier AI Grand Challenge. The goal is to build, within a year, an open source model with more than 400 billion parameters covering all 24 official languages of the European Union, trained using up to 2.5% of EuroHPC's computing capacity for twelve months. The difference from the previous two cases lies in the process: AI Act compliance and the terms of the open release are part of the brief from day one, not a clause added later under public pressure. ## What changes, in practice, for anyone evaluating an open model today - "Open weight" is no longer a guarantee on its own: before choosing a model, verify whether the weights and license are already published — not merely announced — at the time you evaluate it. - Read the full license text, explicitly looking for geographic use restrictions: the MiniMax H3 precedent shows that an "open" model can still exclude Switzerland, the EU, the UK or the US from local deployment. - At equal capability, prefer models with a permissive license published from day one (like MIT or Apache 2.0) over models with weights "coming soon", especially if the use case touches sensitive data or production systems. - Keep an eye on EUROPA over the coming months: an open model designed to be AI Act-compliant from the ground up could become the simplest option to adopt in a regulated European context, once released. This is exactly the check we run for our clients before building an architecture on an open model: it isn't enough for a vendor to use the word "open" in a press release. You need to verify, license in hand, where you can actually run it and under what conditions — before it becomes a compliance problem in production, not after. ### Sources - [European Commission — official announcement on the selection of the EUROPA consortium](https://digital-strategy.ec.europa.eu/en/news/commission-selects-europa-consortium-winner-frontier-ai-grand-challenge-project-build-european-open) - [The New Stack — developer reactions to Qwen3.8-Max: "an API business model wearing an open source jacket"](https://thenewstack.io/alibaba-qwen3-8-max-reactions/) - [Tech Times — MiniMax H3's open weights exclude the US, EU, UK and South Korea from local deployment](https://www.techtimes.com/articles/322904/20260804/minimax-h3-open-weights-exclude-us-eu-uk-korea-local-deployment.htm) - [Slator — EU backs an open source AI model covering 24 European languages](https://slator.com/eu-open-source-ai-model-24-eu-languages/) - [DeepSeek — official DeepSeek-V4 page (V4-Pro and V4-Flash)](https://deepseek.ai/deepseek-v4) - [Qwen — official Qwen3.8-Max announcement](https://qwen.ai/blog?id=qwen3.8) *** - [all articles](/en/markdown.md) - [formatted version](/en/blog/open-weight-o-open-washing-agosto-2026) © 2026 ai.malagoli.me · data hosted in Switzerland · CH ✓ [Privacy Policy](/en/privacy) · [Cookie Policy](/en/cookie-policy)