Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-14

Top developments

Mistral AI Secures €3B at €21B Valuation

On September 8, 2026, French AI lab Mistral AI announced a €3 billion Series D funding round, bringing its valuation to €21 billion. The round was led by Samsung, Scaleup Europe, and PSG Equity, signaling strong industrial and sovereign support for European open-weight models. This capital injection is expected to accelerate the development of future open-weight releases and strengthen Mistral's "sovereign AI" positioning against US-based competitors.

Japan Releases LLM-jp-4-VL 9B Multimodal Model

The National Institute of Informatics (NII) in Japan released LLM-jp-4-VL 9B, a new vision-language model built upon the earlier LLM-jp-4 architecture. Published in late August/early September 2026, this Apache 2.0 licensed model aims to provide high-quality Japanese multimodal capabilities for local deployment. It complements the recent release of the text-only LLM-jp-4 33B, offering a smaller, more accessible option for developers needing visual understanding in Japanese contexts.

Qwen3.8 Dominates Hugging Face Trends

Data from early September 2026 indicates that Alibaba's Qwen family continues to dominate the open-weight landscape on Hugging Face. The unsloth/Qwen3.8-27B-GGUF model has surpassed 10 million downloads, reflecting massive demand for efficient, locally runnable models. The ecosystem is bifurcating between massive proprietary-style open weights (like Qwen 3.8 and GLM 5.3) and a vibrant community of quantizers creating GGUF and MLX formats for Mac and local hardware.

Local view

France: Sovereignty and Valuation French media highlights Mistral's new valuation as a validation of the "open-weight" strategy for secure enterprise AI. Le Big Data notes that the €21 billion valuation proves open-weight models are viable for secure, sovereign data processing, a key selling point for European CIOs. However, some outlets like Cafetech remain skeptical, questioning whether Mistral's storytelling has outpaced its actual competitive edge against US giants.

Japan: Local Language Models Gain Momentum Japanese tech media is focusing on the practical utility of newly released domestic models. ITmedia analyzes the economic rationale behind companies releasing expensive-to-train models for free, noting that state-backed initiatives like the NII's LLM-jp series are crucial for maintaining linguistic sovereignty in AI. The release of LLM-jp-4-VL 9B is seen as a step toward more robust local multimodal applications that respect Japanese cultural and linguistic nuances.

Context & numbers

• Mistral Valuation: €21 billion post-money valuation following the €3 billion Series D round.
• Qwen Downloads: The unsloth/Qwen3.8-27B-GGUF repository exceeded 10 million downloads in early September 2026.
• LLM-jp-4 Specs: The newly released VL model is 9 billion parameters, while the base text model (released Aug 18) is ~33 billion parameters, both under Apache 2.0 licenses,.
• GGUF Dominance: GGUF builds of large models like DeepSeek-V4-Flash (~284B params) and Kimi-K3 (~2.8T params) are now standard in the Hugging Face ecosystem, facilitated by the ggml team's integration with Hugging Face resources.

On the radar

• Mistral's Next Release: With fresh capital, industry watchers anticipate a new flagship open-weight release from Mistral AI in Q4 2026 to maintain its lead over Gemma and Qwen in the Western market.
• Korean Security Models: A consortium led by Naver Cloud is developing 700-billion parameter security-specialized AI models based on HyperCLOVA X and EXAONE, with results expected later this year.