Reflection AI

Artificial Intelligence · Open-Weight Models · Developer Tools

Launch

Reflection AI launches Beam, a 501B-parameter open model to counter China

It claims 3-4x less inference compute than rivals, undercutting the economics behind today's dominant open-source models.

• Reflection AI released Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, built for coding, reasoning and agentic tasks; full weights follow later this month.
• Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, builds AI systems that automate software development.
• Beam was pretrained on 23.8 trillion tokens and refined with high-compute reinforcement learning that generated over 100 million rollouts across 10,500 Nvidia GB300 GPUs over four weeks.
• Reflection says Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while needing three to four times less inference compute, though its own figures show Moonshot's Kimi K3 still ahead on raw capability.
• Reflection closed its latest funding round at a $25 billion pre-money valuation in June, backed by Nvidia, Sequoia and Citigroup, and signed a $6.3 billion compute deal with SpaceX's Colossus 2 data center.
• CEO Misha Laskin told Semafor that Reflection is already training its next model, which he described as 'much more' powerful and performant than Beam.
• Beam matters because Chinese open models have become the default for cost-conscious developers worldwide, and a credible Western alternative could reduce reliance on them for governments and enterprises building sovereign AI systems.