"I don't want an oligopoly": New open-weight AI models mount a comeback against China

A new crop of powerful, Western open-weight AI systems are mounting a comeback, with fresh models from Reflection and Mistral narrowing China's recent lead.Why it matters: These models and others will give businesses and governments alternatives to the closed ecosystems of OpenAI and Anthropic, as well as to Chinese options they may have eschewed for security reasons.But putting powerful model weights into the world also makes their safeguards easier to remove, a concern that's growing as models become more capable in areas like cybersecurity.What they're saying: Both companies are arguing that raw intelligence isn't the only way to measure the AI race, and they present the notion of user control as a core opportunity and selling point."Once you're spending that amount on intelligence, you want to move from renting it to owning it yourself," Reflection CEO Misha Laskin told Axios. "That's kind of where open source is very powerful because it is customizable at every level."Mistral VP of science Pierre Stock said the themes he repeatedly hears from customers include control over data and intellectual property, business continuity, the ability to customize models and cost. "I don't want to live in the future in which any oligopoly controls closed access to this type of intelligence," Stock said in an interview.Zoom in: Reflection on Monday introduced Beam, a 501-billion-parameter mixture-of-experts model, with 23 billion parameters active at a time.Reflection says Beam is competitive with China's GLM-5.2 and approaches Alibaba's Qwen 3.8-Max on some coding and agentic tasks. Its own published benchmarks show a mixed picture: Beam beats or roughly matches the Chinese models on some tests while trailing them on others.Reflection says Beam can achieve reasoning performance comparable to GLM-5.2 with three to four times less computing firepower — an important consideration because the cost of running AI models is now as important to many customers as raw performance.Reflection plans to release Beam's weights later this month, along with tools for running, evaluating and fine-tuning it.Meanwhile, Mistral said Tuesday it is finishing work on a new flagship model, Mistral Large 4, aka "Le Chonk." It's a 1-trillion-parameter multimodal model with 49 billion active parameters, trained on 4,000 Nvidia Grace Blackwell GPUs over two months in Mistral's own European data centers.Stock told Axios the company believes Le Chonk is the world's best open-weight model and said it can outperform closed models on certain tasks.But Stock acknowledged that Mistral hasn't yet caught the leading closed models. "We're not there yet on the frontier," he said.Mistral is initially making the model available through a moderated API, with a version with fewer restrictions and broader cybersecurity capabilities being shared with select partners for testing. It plans to release the weights Oct. 27 after further reinforcement learning and safety testing.The big picture: Chinese com