New open-weight systems from Reflection and Mistral — Beam and Mistral Large 4 ("Le Chonk") — are presenting renewed Western competition to recent Chinese advances in open-weight AI, narrowing what had been a growing gap.
Why this matters
These models provide businesses and governments alternatives to the closed ecosystems of companies like OpenAI and Anthropic, and to Chinese options that some organizations avoid for security reasons. At the same time, publishing powerful model weights makes it easier to remove safeguards, raising concerns as models improve at sensitive tasks such as cybersecurity.
What the companies say
Both firms emphasize that raw capability is not the only metric in the AI race and promote user control as a selling point. Reflection CEO Misha Laskin told Axios: "Once you're spending that amount on intelligence, you want to move from renting it to owning it yourself. That's kind of where open source is very powerful because it is customizable at every level." Pierre Stock, Mistral's vice president of science, listed control over data and intellectual property, business continuity, customization, and cost as recurring customer priorities. "I don't want to live in the future in which any oligopoly controls closed access to this type of intelligence," Stock said.
Reflection's Beam
On Monday Reflection unveiled Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters 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 published benchmarks show a mixed picture: Beam outperforms or roughly matches the Chinese models on some tests and trails them on others.
Reflection asserts Beam can achieve reasoning performance comparable to GLM-5.2 using three to four times less compute, an important point because operating costs are now as significant to many customers as raw performance. The company plans to release Beam's weights later this month, alongside tools for running, evaluating, and fine-tuning the model.
Mistral's Mistral Large 4 ("Le Chonk")
Mistral said on Tuesday it is finalizing its flagship model, Mistral Large 4, nicknamed "Le Chonk." The model has 1 trillion parameters and 49 billion active parameters, and was trained over two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's European data centers.
Stock told Axios that the company believes Le Chonk is the world's best open-weight model and can outperform closed models on certain tasks, but he also acknowledged Mistral "is not there yet on the frontier" compared with top closed models.
Mistral will initially make the model available via a moderated API, while a less-restricted version with broader cybersecurity capabilities is being shared with selected partners for testing. The company plans to release the model weights on Oct. 27 after additional reinforcement learning and safety testing.
Bigger picture and trade-offs
Chinese firms including Alibaba, Z.ai, Moonshot and DeepSeek have driven much of the recent momentum in open-weight AI. The arrival of competitive models from Reflection and Mistral suggests Western labs may be closing a gap that had begun to look structural.
However, the same trait that gives customers control over an open-weight model — public access to weights — reduces developers' control over how the model is used. Once weights are public, users can modify the model, including attempts to remove safeguards the creators added. This is especially significant as models become more capable in cybersecurity and other potentially harmful domains.
Openness as a safety argument
Mistral and Reflection counter that openness can improve safety. Laskin argued that a larger ecosystem able to inspect models and hunt for vulnerabilities is safer than concentrating that work among a few hundred researchers inside closed labs, though he said capability-based safety thresholds should determine whether any model — open or closed — should be released.
Stock acknowledged the trade-off in releasing weights but said making capable models available can accelerate cyber defenses and allow external researchers to audit the technology. Mistral's staged release is intended to give extra time for testing. Stock added that open-weight releases help defenders obtain adequate tools: "If we open weight as many models as possible, including ML4 — which is among the best models in the world on cyber — then you accelerate the defense part way more."
Conclusion
Reflection and Mistral's new models increase Western options in the open-weight landscape, appealing to customers seeking control, customization and cost-efficiency. At the same time, releasing model weights raises safety concerns about misuse and the removal of safeguards, meaning the timing and manner of releases will remain a key policy and technical debate.
