News chronological

The article identifies Alibaba, Z.ai, Moonshot, and DeepSeek as Chinese companies that have driven much of the recent momentum in the development of open-weight AI models.

The primary concern is that making model weights publicly available allows users to modify the models and potentially remove safeguards, which is a significant risk as models become more capable in areas like cybersecurity. In response, Mistral and Reflection argue that openness can actually improve safety by allowing a larger ecosystem of researchers to inspect models and hunt for vulnerabilities, rather than concentrating that work within a few closed labs. They also argue that open-weight models provide those defending against cyberattacks with better tools to accelerate cyber defenses.

The companies argue that open-weight models provide users with ownership, customization, and control over data and intellectual property, rather than "renting" intelligence from an oligopoly. Reflection CEO Misha Laskin emphasized that open source is powerful because it is customizable at every level. Mistral's VP of science, Pierre Stock, noted that customers prioritize business continuity, cost, and the ability to customize models, stating that he does not want a future where an oligopoly controls closed access to such intelligence.

Mistral Large 4 is a 1-trillion-parameter multimodal model with 49 billion active parameters. It was trained over a period of two months using 4,000 Nvidia Grace Blackwell GPUs located in Mistral's own European data centers. Mistral intends to release the weights for this model on October 27, following a period of reinforcement learning and safety testing. Before the full release, it is being made available through a moderated API and shared with select partners for testing.

Reflection introduced Beam, a 501-billion-parameter mixture-of-experts model with 23 billion parameters active at a time. Reflection claims that Beam is competitive with China's GLM-5.2 and approaches Alibaba's Qwen 3.8-Max on specific coding and agentic tasks. Furthermore, Reflection states that Beam can achieve reasoning performance comparable to GLM-5.2 while requiring three to four times less computing firepower, which addresses concerns regarding the cost of running AI models.