China's DeepSeek, peers launch 16 AI models in month despite Anthropic warning

A fresh wave of Western open-weight AI systems is challenging China's dominance, as Reflection and Mistral unveil cutting-edge models. These models offer businesses and governments alternatives to closed ecosystems like OpenAI and Anthropic, and open-source options they may have avoided due to security concerns. However, putting powerful model weights into the world also raises concerns about the ease with which safeguards can be bypassed, especially as models grow more capable in areas like cybersecurity.

Reflection CEO Misha Laskin argues that raw intelligence is not the sole measure of AI prowess, emphasizing user control as a core advantage of open-source models. Mistral VP of science Pierre Stock highlights user control over data, intellectual property, business continuity, customization, and cost as key selling points. Both companies believe openness can enhance safety by allowing a broader ecosystem to scrutinize models and identify vulnerabilities, an advantage over closed labs with limited resources.

Reflection recently launched Beam, a 501-billion-parameter mixture-of-experts model with 23 active parameters, challenging China's GLM-5.2 and rivaling Alibaba's Qwen 3.8-Max on certain tasks. While Beam matches or surpasses some Chinese models in benchmarks, its performance varies across tests. The company plans to release Beam's weights and tools for running, evaluating, and fine-tuning it later this month.

Mistral, meanwhile, is preparing to unveil Le Chonk, a 1-trillion-parameter multimodal model with 49 active parameters. Trained on Mistral's European data centers, the model aims to outperform closed models on specific tasks. Mistral is initially offering Le Chonk through a moderated API, with a more permissive version for select partners after additional safety testing. The company acknowledges the trade-off in releasing weights but believes open models can accelerate cyber defenses and facilitate outside audits.