Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-16
Salesforce has launched "Koa," its first in-house reasoning model built on Nvidia's open-weight Nemotron architecture, signaling a shift toward enterprise-specific fine-tuning of open foundations. Meanwhile, Mistral AI secured a $3.5 billion funding round to bolster its European sovereign AI infrastructure, while Kakao in Korea completed training on its massive 155B-parameter "Kana" model. These developments highlight a growing trend of major tech giants and regional leaders leveraging open-weight models for proprietary, cost-efficient, and locally compliant AI solutions.
Open Weights Outside China: Llama, Mistral, Gemma — 2026-09-16
Top developments
Salesforce Unveils "Koa" Built on Nvidia Nemotron
At the recent Dreamforce conference, Salesforce introduced "Koa," its first proprietary reasoning model, which is fundamentally based on Nvidia’s open-weight Nemotron architecture. This move demonstrates how large enterprises are increasingly bypassing closed frontier models in favor of fine-tuning open-weight bases like Nemotron for specific enterprise workflows, prioritizing data sovereignty and lower inference costs. The release underscores Nvidia’s strategy to make Nemotron a standard foundation for enterprise-grade AI applications outside of pure research contexts.
Mistral AI Raises $3.5 Billion for Sovereign Infrastructure
Mistral AI announced a $3.5 billion fundraising round to accelerate its strategy of providing European sovereign AI alternatives to American and Chinese rivals. The capital will support the expansion of its open-weight model ecosystem, including the widely used Apache-licensed Mistral Large and Small variants, while enhancing its regional endpoints to ensure data residency compliance within the EU. This significant injection of capital reinforces the viability of open-weight models as a commercial product line when coupled with secure, localized infrastructure.
Kakao Completes Training on 155B-Parameter "Kana" Model
South Korean tech giant Kakao has reportedly completed training on "Kana," a new large language model with 155 billion parameters, marking a significant leap from its previous 30B-parameter versions. While details on the exact license for the weights remain under wraps, the development signals Korea’s intensified push to reduce dependence on external AI providers by building robust, locally hosted open-weight or semi-open alternatives. Kakao plans to deploy Kana across various service sizes (0.9B to 155B) to optimize serving costs and response quality for its vast domestic user base.
InternLM Releases Atria Dawn Preview with MIT License
The InternLM team quietly released "Atria Dawn Preview," a 744-billion-parameter Mixture-of-Experts (MoE) agent model, under the permissive MIT license on Hugging Face. This release is notable for its 1 million token context window and its focus on agentic capabilities, offering developers a powerful, unrestricted foundation for complex autonomous tasks. The MIT licensing allows for seamless commercial integration, contrasting with more restrictive licenses often seen in other large-scale open-weight releases.
Local view
In France, Le Big Data analyzed Mistral AI’s valuation at €21 billion, arguing that the company’s success proves open-weight models are the future of secure enterprise AI. The outlet highlights that CIOs are increasingly preferring open-weight solutions for data security and control, provided they are backed by strong local infrastructure like Mistral’s sovereign endpoints.
In Japan, Forbes Japan published a critical piece questioning whether "open-weight" truly equates to "open-source," urging enterprises to scrutinize the costs, control rights, and digital sovereignty implications of using these models. The article suggests that while weights are public, the ecosystem dependencies and potential hidden costs require careful governance.
Context & numbers
- Enterprise Adoption: Over 40% of enterprise AI workloads now include a local inference component, driven by the need for privacy and cost control.
- Download Growth: Hugging Face reports that downloads of quantized model weights (such as GGUF formats for local execution) have grown by 320% year-over-year, indicating a massive shift toward local deployment of open-weight models like Llama, Gemma, and Mistral.
- Licensing Trends: Apache 2.0 and MIT licenses remain dominant for enterprise-friendly open-weight models, with Qwen3 and InternLM’s recent releases adhering to these standards to maximize commercial uptake.
On the radar
- Nemotron Ecosystem Expansion: Following Salesforce’s Koa launch, watch for other enterprise SaaS providers announcing similar "built-on-Nemotron" models as Nvidia aggressively markets its open-weight family.
- Korean Open Weights Release: Industry observers are waiting for Kakao to officially publish the weights or API access details for the newly trained 155B Kana model, which could disrupt the local Korean AI market.
- EU AI Act Compliance: With the AI Act fully in effect, monitor how Mistral and other European labs adjust their open-weight licensing terms to comply with transparency requirements for high-risk applications.
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
