Unlike closed models from OpenAI and Anthropic, open-weight systems allow users to inspect, customize, and run models themselves (Representational image)
Western AI companies are mounting a comeback in the race for powerful open-weight models, challenging the recent momentum of Chinese firms.
Reflection and Mistral are developing new systems that could give businesses and governments greater control over advanced AI. Unlike closed models from OpenAI and Anthropic, open-weight systems allow users to inspect, customize, and run models themselves.
New models enter the race
Reflection recently unveiled Beam, a 501-billion-parameter mixture-of-experts model, with 23 billion parameters active at a time. The company says Beam can compete with leading Chinese models and requires less computing power.
The text-only model is designed for reasoning, coding and agentic tasks, with a focus on lowering inference costs and reducing processing time.
Reflection plans to release Beam’s full technical details this month.
Meanwhile, Mistral is preparing Mistral Large 4, called “Le Chonk”—a 1-trillion-parameter multimodal model trained on 4,000 Nvidia GPUs. The France-based company plans to release its model weights on October 27.
The model is currently available through Mistral’s API during a limited preview.
If the new models live up to their claims, they could shake up the AI industry.
They could provide an alternative to Chinese AI models that some organizations avoid because of security concerns.
OpenAI and Anthropic may face fresh competition as security concerns mount and the companies position themselves for potential public offerings.
However, open-weight AI also creates safety concerns. Users can modify models and potentially remove safeguards. As AI becomes more capable in cybersecurity, those risks are growing.
A changing AI race
Both companies argue that openness can strengthen safety by allowing researchers to inspect models and improve defenses.