Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-11
Mistral AI has closed a massive €3 billion funding round at a €21 billion valuation, signaling a strategic pivot toward sovereign AI infrastructure while maintaining its open-weight leadership. Meanwhile, NVIDIA quietly open-sourced the Nemotron-3-Labs-Ultra-Math-RL checkpoint, and Kakao completed training on its new 155B-parameter "Kakana" model. The ecosystem continues to consolidate around local inference capabilities, with recent data highlighting significant disparities in GGUF download volumes between major model families.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-11
Top developments
Mistral AI Secures €3B to Expand Sovereign Open-Weight Infrastructure
On September 8, 2026, Mistral AI announced a Series D round raising €3 billion at a €21 billion valuation, led by Samsung, Scaleup Europe, and PSG Equity. The French lab intends to direct this capital toward frontier research, compute, and infrastructure, reinforcing its position as a European alternative for secure, customizable AI. This move addresses growing market concerns about whether open-weight models can compete with proprietary giants like OpenAI and Anthropic without substantial infrastructure backing.
NVIDIA Releases Nemotron-3-Labs-Ultra-Math-RL Checkpoint
NVIDIA quietly open-sourced the Nemotron-3-Labs-Ultra-Math-RL checkpoint, which was part of its gold-level performance run in the IMO 2026 competition. Released approximately two days ago, this model provides the reinforcement learning (RL) recipe and specifications behind NVIDIA's high-scoring mathematical reasoning capabilities. This release adds to the Nemotron family, offering developers access to specialized reasoning weights that complement the previously released Nemotron 3.5 Lightning models.
Kakao Completes Training of 155B-Parameter "Kakana" Model
Korean tech giant Kakao has effectively completed training on its new "Kakana" LLM, which boasts 155 billion parameters—more than five times the size of its previous 30B model. The company is building a suite of models ranging from 0.9B to 155B parameters to optimize serving costs and reduce dependency on external AI providers. This development marks a significant step for Korean sovereign AI efforts, positioning Kakao alongside other domestic players like LG AI Research and Naver in the open-weight space.
"Model Fatigue" Emerges as Labs Race to Release Updates
A CNBC report from September 6 highlighted growing "model fatigue" among users and enterprises as Meta, Google, OpenAI, and Anthropic all released updates in a single week. The frenetic pace of releases, including new versions of Gemini and Claude, has made it difficult for developers to keep up with benchmark changes and integration requirements. This sentiment underscores the challenge for open-weight models like Llama and Gemma to maintain ecosystem engagement when proprietary labs dominate the news cycle with frequent, incremental improvements.
Local view
Japan: ITmedia Analyzes Economics of Free Open Weights
Japanese outlet ITmedia published an analysis on September 10 exploring how vendors sustain the "local LLM boom" by offering expensive-to-train models for free. The article examines the strategic motivations of vendors and governments in releasing open-weight models, noting that despite high development costs, these releases serve broader ecosystem and sovereignty goals.
Korea: Naver's HyperCLOVA X Excluded from National "Everyone's AI" Consortium
South Korean media reported that Naver's HyperCLOVA X was notably absent from the list of domestic AI models selected by the three consortiua participating in the government's "Everyone's AI" (Moodeuui AI) initiative. This exclusion raises questions about the competitive dynamics within Korea's national AI strategy, even as other domestic models like LG's EXAONE and Kakao's new offerings gain traction.
Context & numbers
Download Volumes Reveal Qwen Dominance Over Llama and Gemma
Recent data from Hugging Face indicates that Qwen models are leading the local inference market with 39.6 million GGUF downloads per month. This figure is nearly twice that of Google's Gemma (20.8 million downloads) and more than five times that of Meta's Llama (7.5 million downloads). Despite this gap, Llama-derived GGUF repositories slightly outnumber Qwen's, suggesting that the disparity is driven by community adoption rather than supply.
License Audit Highlights Apache 2.0 Preference
A comprehensive audit of open-weight licenses notes that Apache 2.0 and MIT remain the cleanest options for commercial builders, while custom licenses introduce complexities regarding user caps and geography. Models like Qwen3 235B-A22B are highlighted as safe enterprise picks due to their explicit Apache 2.0 licensing, contrasting with the more restrictive terms sometimes associated with newer frontier releases.
On the radar
- Hugging Face Acquisition Rumors: Reports suggest Nvidia is acquiring the open-source platform Hugging Face, though this remains unconfirmed by both parties as of early September.
- LLM-jp-4-VL 9B Release: The Japanese LLM-jp project released its 9B-parameter vision-language model, LLM-jp-4-VL 9B, extending the capabilities of the domestic Apache 2.0 licensed family.
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
