Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-06
Mistral AI has sparked debate by hosting the Chinese Z.ai GLM-5.2 model on its European infrastructure, signaling a shift in European AI sovereignty strategies. Meanwhile, Japan’s NII released LLM-jp-4-VL 9B, a new multimodal open-weight model, while Hugging Face data reveals Qwen's dominant lead in local GGUF downloads over Llama and Gemma.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-06
Top developments
Mistral AI hosts Chinese Z.ai GLM-5.2 on European servers
Since August 11, 2026, Mistral AI has offered access to GLM-5.2, a large language model developed by the Chinese laboratory Z.ai, through its own API and Vibe platform hosted on European servers. This move has triggered significant discussion within the French tech community, with some critics viewing it as a departure from strict European sovereignty ideals, while others praise the pragmatic approach to providing high-performance open-weight alternatives regardless of origin.
Japan releases LLM-jp-4-VL 9B multimodal model
On September 1, 2026, the National Institute of Informatics (NII) in Japan released LLM-jp-4-VL 9B, a new vision-language model derived from the LLM-jp-4 series. This 9-billion parameter model expands the capabilities of the Japanese open-weight ecosystem into multimodal tasks, following the earlier release of the text-only LLM-jp-4 33B in August. The release underscores the continued investment in sovereign, Apache-licensed open models for Japanese language processing.
Qwen dominates local inference downloads over Llama and Gemma
Hugging Face's "State of Open Models" report highlights a stark disparity in local adoption rates as of late August 2026. Qwen models have achieved approximately 39.6 million GGUF downloads per month, nearly double the 20.8 million for Google Gemma and more than five times the 7.5 million for Meta’s Llama. Despite Llama-derived repositories slightly outnumbering Qwen's in total count, the download volume indicates a strong community preference for Qwen in local inference environments.
Local view
France: French media outlets Frandroid and Rolling Stone France are actively debating Mistral AI's decision to host Z.ai's GLM-5.2. While some voices in the community express concern over relying on Chinese models, others argue that providing access to top-tier open weights via European infrastructure is a practical step toward "distributed intelligence" rather than strict isolationism.
Japan: Japanese tech media, including AI Revolution and the official NII blog, are focusing on the rapid iteration of domestic open models. The release of LLM-jp-4-VL 9B is being highlighted as a key milestone for Japanese multimodal AI, with detailed analyses of its performance against global benchmarks like GPT-4o in specific Japanese contexts.
Context & numbers
- Download Volumes: Qwen leads local GGUF downloads at ~39.6 million/month, compared to Gemma at ~20.8 million and Llama at ~7.5 million.
- Model Sizes: The trend toward larger open-weight models continues, with GGUF builds now available for DeepSeek-V4-Flash (~284B parameters) and Kimi-K3 (~2.8 trillion parameters).
- Licensing: A recent audit of 2026 open-weight models notes that while many use standard licenses like Apache 2.0, several vendors are reproducing the "user-threshold" license pattern previously seen in Meta’s Llama 4 releases.
On the radar
- Upcoming Model Releases: The AI Flash Report tracker notes that new models are arriving roughly every two days, with recent additions including Microsoft BitNet embeddings and NVIDIA Nemotron variants.
- OWASP Standards: The OWASP GenAI Security Project formally unveiled its 2026 Top 10 for LLM Applications and the new Agent Control Standard (ACS) on September 1-2, 2026, which will likely influence how open-weight models are deployed in enterprise environments.
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
