AI This Week: GLM-5, Qwen3.5, and the Chinese open-source wave

China dropped two frontier-class open-source models in five days. Europe started putting real money behind AI sovereignty. The gap between open and closed models got a lot smaller. Busy week.

Top stories this week

GLM-5: China's first publicly listed AI company ships a frontier model under MIT

Zhipu AI released GLM-5 on February 11th. The numbers: 744 billion parameters in a mixture-of-experts architecture with 40 billion active per token. It scores 77.8% on SWE-bench Verified, putting it ahead of Gemini 3 Pro (76.2%) and within reach of Claude Opus 4.6 (80.9%).

Two things make this release stand out beyond the benchmarks.

The model was trained entirely on Huawei Ascend chips using the MindSpore framework. Zero NVIDIA hardware. For anyone tracking the US-China chip export controls, this is the clearest signal yet that sanctions haven't stopped Chinese AI progress. They've redirected it.

The license is MIT. Not "open with restrictions" or "research only." MIT. You can run it, modify it, sell products built on it. At roughly $0.80 per million input tokens on OpenRouter, it costs about six times less than Opus 4.6. For teams that need strong coding performance without sending data to US clouds, this changes the math entirely.

Zhipu's stock surged 34% in Hong Kong after the announcement. They're China's first publicly listed pure-play AI company, and GLM-5 is their proof that being public and being competitive aren't mutually exclusive.

Qwen3.5: Alibaba answers with 201 languages and aggressive pricing

Five days later, Alibaba released Qwen3.5 on February 16th. Another MoE model: 397 billion total parameters, 17 billion active per token, trained on text and images simultaneously through early fusion. The Apache 2.0 license makes it fully open-weight.

The language coverage is striking. Qwen3.5 supports 201 languages and dialects, up from 119 in the previous version. For anyone building AI products outside the English-speaking world, that kind of multilingual breadth used to require proprietary models.

On cost, Alibaba claims 60% lower pricing and 8x throughput compared to their previous flagship Qwen3-Max. The 1M-token context window costs roughly $0.18 per million input tokens. That puts it in a completely different pricing tier.

I've been watching the Qwen releases closely because the MoE architecture (huge model, tiny active footprint) is exactly what makes self-hosting viable. 17 billion active parameters means you can run this on hardware that wouldn't touch a dense 397B model. If you want frontier performance without leaving your own servers, this is the kind of model that gets you there.

Mistral puts 1.2 billion euros into Swedish AI infrastructure

On February 11th, Mistral announced a 1.2 billion euro partnership with EcoDataCenter to build AI compute infrastructure in Borlange, Sweden. The facility will host NVIDIA Vera Rubin GPUs and run on renewable energy. Operational target: 2027.

This is Mistral's first infrastructure investment outside France. The explicit goal: a fully European AI stack designed, built, and operated across the entire value chain, with data processed and stored locally in Europe.

For context, the EU just committed to 200 billion euros in AI infrastructure as part of its sovereignty push. Mistral is positioning itself as the model provider for that stack. Given their Apache 2.0 licensing on models like Mistral Large 3, this is a credible play: European governments need compute plus models they can actually audit and control.

Four EU governments commit to European-made AI for public services

At a February 12th summit in Belgium, Germany, Poland, Spain, and the Netherlands committed to deploying European-made AI in public administration. This aligns with the EU's new "Buy European" policy, which prioritizes European firms in strategic sectors like AI.

Each country is backing its own initiative: Germany with SOOFI (targeted for public sector deployment by mid-2026), Poland with PLLuM (a Polish-language model now expanding into administrative tools), Spain with Alia (running on the MareNostrum 5 supercomputer for multilingual public services), and the Netherlands with GPT-NL (focused on Dutch-language health, education, and government applications).

These aren't vague policy announcements. They're funded programs with deployment timelines. When four of the largest EU member states simultaneously commit to sovereign AI, it creates a procurement pipeline that benefits every European AI builder, Mistral included.

What does this actually mean?

The week of February 9-15 was the week open-source AI reached credible frontier performance. GLM-5 and Qwen3.5 aren't "good for open-source." They're good, full stop. Both compete with the best proprietary models on coding and reasoning benchmarks, at a fraction of the cost, with permissive licenses.

For practitioners, this shifts the default assumption. A year ago, you chose open-source if you couldn't afford proprietary. Now you choose open-source because the performance is there and you keep control of everything else. The closed-model tax is getting harder to justify.

Europe noticed. Mistral's infrastructure play and the EU sovereign AI commitments aren't happening in a vacuum. When viable open-source models exist at frontier level, the argument for European AI sovereignty stops being theoretical. You can actually build it.

What to watch next week

• Claude Sonnet 4.6 is expected this week: Anthropic's next Sonnet release should land, potentially bringing Opus-level performance at Sonnet pricing. If it delivers, the cost-performance landscape shifts again.
• Gemini 3.1 Pro is rumored: Google may follow with its own reasoning-focused release, potentially doubling Gemini 3 Pro's performance.
• The February model rush continues: we're already at four frontier releases this month (Opus 4.6, GLM-5, Qwen3.5, plus earlier drops). By month's end, the count could hit seven. Unprecedented density.