Reflection Launches Beam Open-Weight AI Model to Challenge Chinese Rivals
Reflection AI officially unveils Beam, the startup's first frontier open-weight AI model. The two-year-old Brooklyn company claims Beam matches leading Chinese open models like DeepSeek, Qwen, and Z.ai on advanced reasoning benchmarks while using dramatically less compute, positioning itself as a Western answer to those rivals.
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, pretrained on 23.8 trillion tokens with a 1 million token context window. Trained with high-compute reinforcement learning, it targets reasoning, coding, and agentic tasks. Reflection says Beam scores on par with Z.ai's GLM-5.2 and outperforms leading Western open models while using 3-4x less inference compute, though the claims remain independently unverified.
Reflection competes against closed labs like Anthropic and OpenAI, Chinese open-model developers, and Western players including Mistral, Meta, Cohere, and Mira Murati's Thinking Machines Lab, whose Inkling model it beats on four coding benchmarks. Founded in 2024 by former Google DeepMind researchers, the company has raised roughly $4.7 billion from Nvidia, Sequoia, and Lightspeed at a $25 billion pre-money valuation, and has signed over $7 billion in compute deals with SpaceX and Nebius to secure Nvidia GB300 chips.