Open Weights Outside China: Llama, Mistral, Gemma — 2026-10-15
Mistral AI’s release of its 1-trillion-parameter model "Le Chonk" in API preview has intensified the race for open-weight dominance outside China, with full weights scheduled for late October. Meanwhile, Google’s Gemma 4 family continues to dominate local deployment trends on Hugging Face, and Naver is expanding its sovereign AI footprint in Southeast Asia.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-10-15
Top developments
Mistral Large 4 ("Le Chonk") Enters API Preview
Mistral AI launched the API preview of Mistral Large 4, a 1.05 trillion parameter Mixture-of-Experts model nicknamed "Le Chonk," on October 6, 2026. The company claims this model rivals the best open systems from China, aiming to reduce the performance gap with proprietary US models. Full open weights are scheduled for public release on October 31, 2026, marking a significant attempt by a Western lab to compete at the frontier tier of open-weight AI.
Gemma 4 Dominates Local Model Trends
Google’s Gemma 4 family, particularly the 2B and 31B variants, remains highly active in the community ecosystem as of early October 2026. Recent Hugging Face trending reports highlight that Gemma-based models are frequently adapted into GGUF formats for local inference, with a 2B embedding variant seeing rapid adoption for semantic search and RAG indexing. The family’s Apache 2.0 licensing continues to make it a preferred choice for commercial fine-tuning and enterprise integration outside of China.
Naver Expands Sovereign AI Abroad While Retiring HyperCLOVA X Domestically
South Korean tech giant Naver is shifting its strategy from domestic general-purpose LLMs to localized sovereign AI solutions abroad. As reported on October 8, 2026, Naver completed a paid project for a Thai client based on its HyperCLOVA X technology, while officially retiring the consumer-facing HyperCLOVA X service in South Korea to integrate AI more deeply into its core search and shopping platforms. This move highlights a trend among Asian labs to monetize model weights through customized, local-language deployments rather than competing directly with global frontier APIs.
Nemotron 3 Diarization Model Released
Nvidia released Nemotron 3 Diarization, an open-weight streaming speaker diarization model, as part of its October 2026 update cycle. This model is designed to determine "who spoke when" in real-world audio scenarios, expanding the utility of the Nemotron family beyond pure text generation into multimodal audio processing. It joins the broader ecosystem of specialized open-weight models that complement larger language models like Llama and Mistral.
Local view
France: French media outlets like Le Monde and IntelliNews are closely following Mistral’s "Le Chonk" launch, framing it as a critical effort to prove Europe’s viability in frontier AI. IntelliNews notes that while open models generally lag proprietary ones by six months, Mistral’s release is seen as a necessary step to maintain sovereignty and reduce dependency on Chinese or American closed systems.
Japan: Japanese tech blogs are analyzing the release of Liquid AI’s LFM2.5-24B-A2B (released Oct 13, 2026), noting its efficiency for on-device agents with only 2.45B active parameters out of 24B total. This aligns with the Japanese interest in "local-first" AI, where models like Sakana AI’s orchestration tools and NII’s LLM-jp-4 are optimized for specific hardware constraints rather than raw scale.
Context & numbers
The global Large Language Model market is projected to reach USD 148.8 billion by 2036, with South Korea showing a CAGR of 31.5% and Japan 32.4%, indicating strong regional growth in AI infrastructure and model development.
In the open-weight sector, security concerns remain prominent; recent audits of LiteLLM (v1.77.7) uncovered high-impact vulnerabilities (CVE-2026-35029, CVE-2026-59823), now resolved in v1.83.9. This underscores the need for rigorous security practices when deploying open-weight gateways and models in enterprise environments.
On the radar
- October 31, 2026: Mistral Large 4 ("Le Chonk") open weights are due for public release.
- October 15, 2026: StepFun promised to release open weights for its Step-5-Preview-BF16 model, though as of Oct 8, the Hugging Face repo contained no weights or license.
- Upcoming: Speculation continues around a new "fat but sparse" Mixture-of-Experts family from Mistral, rumored to enter early access later this year.
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
