Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-10
Mistral AI secured a massive €3 billion funding round led by Samsung, valuing the French lab at €21 billion and reinforcing its "sovereign AI" open-weight strategy. Meanwhile, NVIDIA quietly released the Nemotron-3-Labs-Ultra-Math-RL checkpoint following its IMO 2026 success, and TII’s Falcon team launched the K2 Horizon family with six Apache 2.0 models. In South Korea, Kakao completed training its massive 155B parameter "Kanaana" LLM, signaling a shift away from external dependencies for domestic serving costs.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-10
Top developments
Mistral AI Valued at €21 Billion After €3B Raise
Mistral AI closed a Series D funding round of €3 billion on September 8, 2026, led by Samsung, Scaleup Europe, and PSG Equity. This valuation surge to over €21 billion underscores the market's confidence in Mistral's "sovereign AI" approach, which prioritizes open-weight models that enterprises can customize and host locally. The move counters recent analyst doubts about Mistral's competitiveness against US and Chinese giants, reaffirming its commitment to producing frontier-level open models like the upcoming "fat but sparse" MoE family.
NVIDIA Releases Nemotron-3-Labs-Ultra-Math-RL Checkpoint
On September 9, 2026, NVIDIA open-sourced the Nemotron-3-Labs-Ultra-Math-RL checkpoint, the specific reinforcement learning model responsible for its gold-level performance (30/42) at the IMO 2026. This release provides the full specs and training recipe for one of the strongest mathematical reasoning open-weight models available. It strengthens the Nemotron ecosystem's position in specialized reasoning tasks, offering developers a transparent, high-performance alternative to closed-source frontier models for mathematical and logical agents.
TII Launches K2 Horizon Family with Six Apache 2.0 Models
The Technology Innovation Institute (TII), known for its Falcon series, launched the K2 Horizon model family on September 3, 2026. The release includes six fully open models ranging from 0.9B to 375B parameters, all licensed under Apache 2.0. While data and code are promised to follow, the model cards are already live, providing a new competitive option in the open-weight space with clear permissive licensing for commercial and research use.
Kakao Completes Training of 155B Parameter "Kanaana" Model
South Korean tech giant Kakao has effectively completed training its proprietary Large Language Model, "Kanaana," which features 155 billion parameters. Reported by Edaily on September 6, 2026, this model is more than five times larger than Kakao's previous public 30B version. The company aims to reduce reliance on external AI providers and lower serving costs by deploying this in-house model across its services, marking a significant step in Korea's sovereign AI infrastructure.
Local view
France: French media widely covered Mistral's funding as a validation of European digital sovereignty. Le Monde noted that the raise is a direct response to skepticism about Mistral abandoning the frontier model race, emphasizing that the capital will fuel both model development and sovereign infrastructure. Le Big Data highlighted that Mistral's open-weight strategy is now seen as the future of secure enterprise AI in Europe, contrasting it with closed US alternatives.
South Korea: Local coverage focused on the strategic autonomy implied by Kakao's new model. Edaily reported that Kakao is building a suite of models from 0.9B to 155B to handle varying question complexities internally. Additionally, E-Focus pointed out that Naver's HyperCLOVA X was notably absent from the government's "Everyone's AI" consortium list, suggesting shifting dynamics among Korean tech players regarding national AI initiatives.
Context & numbers
- Valuation: Mistral AI's post-money valuation stands at over €21 billion following the €3 billion Series D round.
- Model Size: Kakao's new Kanaana model has 155 billion parameters, compared to the 30 billion parameters of its predecessor.
- Licensing: TII's K2 Horizon models are released under the permissive Apache 2.0 license, covering sizes from 0.9B to 375B parameters.
- Market Sentiment: CNBC reported "model fatigue" setting in due to the frenetic pace of releases from Anthropic, OpenAI, Meta, and Google, alongside NVIDIA's acquisition of Hugging Face.
On the radar
- Mistral MoE Release: Mistral confirmed in early July that a larger "fat but sparse" MoE family is in early access with research partners; a public ship date and parameter count are still pending but expected soon given the fresh capital.
- Nscale Pre-IPO: Reports indicate Nscale is pursuing a $3.5 billion pre-IPO raise, which could significantly impact the compute landscape available for open-weight model fine-tuning and serving.
- LLM-jp Updates: The Japanese NII LLM-jp team recently released LLM-jp-4-VL 9B, a vision-language model, continuing the momentum of Japanese domestic open-weight development.
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
