Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-03
Mistral AI’s decision to host Chinese open-weight models like GLM-5.2 on its European infrastructure has sparked debate about the definition of "sovereign AI," with French media analyzing the strategic pivot. Meanwhile, Japan's National Institute of Informatics (NII) released the multimodal LLM-jp-4-VL 9B, reinforcing the momentum of non-US open-weight ecosystems. The broader landscape sees continued dominance by Apache 2.0 licensed models, with local inference tools like Ollama updating to support these new releases.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-03
Top developments
Mistral AI hosts Chinese GLM-5.2, redefining European sovereignty
Mistral AI has begun hosting GLM-5.2, an open-weight model from Chinese lab Z.ai, on its European API and "Vibe" platform. This move, which started in early August and continues to draw attention in early September, signals a shift where European labs prioritize model performance and ecosystem openness over strict geographic origin for weights. French outlet Frandroid notes that while some community members view this as a betrayal of sovereignty, others see it as a pragmatic strategy to keep the platform competitive by aggregating the best available open weights.
NII releases LLM-jp-4-VL 9B for Japanese multimodal tasks
On August 31, 2026, Japan's National Institute of Informatics (NII) released LLM-jp-4-VL 9B, a vision-language model designed specifically for Japanese contexts. This follows the August 18 release of the text-only LLM-jp-4 33B. The new 9B model is Apache 2.0 licensed, allowing for commercial use, and is positioned as a lightweight, high-performance alternative for local deployment in Japanese enterprises. It builds on the success of the 33B model, which reportedly outperformed gpt-oss-20b in specific Japanese benchmarks.
Ollama v0.33.1 released to support latest open-weight formats
On August 26, 2026, Ollama released version v0.33.1, with a v0.33.2 release candidate following on August 27. This update is critical for the local AI community as it ensures compatibility with the latest quantization formats and model architectures emerging from recent open-weight releases. The update supports the growing trend of running large models locally, which has been accelerated by the "30B parameter" wave seen in August, including models like Nemotron 3.5 Lightning and Qwen3.8-27B.
Local view
France: L'Opinion published an op-ed by David Lacombled on September 1, 2026, discussing "headwinds" for Mistral AI. The piece argues that for Mistral to remain viable against US and Chinese giants, it must open up to foreign funding and solutions, including hosting non-European models. This reflects a growing sentiment in French tech media that strict "sovereign AI" definitions may need to soften to accommodate global open-weight standards.
Japan: ITmedia AI+ highlighted the "heat" around 30B-class open models, noting that within nine days in August, Meta (Muse Glimmer), NVIDIA (Nemotron 3.5 Lightning), and Alibaba (Qwen3.8-27B) all released models in this class. The article emphasizes that Japanese stakeholders are closely watching how these models compare to domestic efforts like LLM-jp, particularly regarding license permissiveness (Apache 2.0 vs. custom licenses) and performance on Japanese-language tasks.
Context & numbers
The Hugging Face "State of Open Models" report indicates that while Qwen leads in GGUF downloads (39.6 million/month), Llama-derived GGUF repositories remain numerous (7.5 million/month). However, the gap is narrowing as newer models like Gemma and Nemotron gain traction. The industry standard remains Apache 2.0 for most new releases, though some vendors still use user-threshold licenses similar to older Llama 4 patterns.
On the radar
- Llama 4 Successor Rumors: While no official date is set, community trackers are monitoring for any updates to Meta's Llama family, especially given the strong competition from Qwen and Nemotron in the 30B+ class.
- GLM-5.3 Open Weights: Following the announcement that GLM-5.3 would be released as an open model, developers are awaiting the actual weight files on Hugging Face, expected to rival Claude Fable 5 and GPT-5.6 Sol in benchmarks.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
