Reflection AI, a startup founded in March 2024 by former Google DeepMind researchers, has introduced Beam, an open-weight model featuring 501 billion parameters. The model utilizes a sparse Mixture-of-Experts architecture, where only 23 billion parameters are active per token. Pretraining involved 23.8 trillion tokens, followed by reinforcement learning using 10.5K NVIDIA GB300 GPUs.
The company claims Beam competes with Chinese models like Z.ai’s GLM-5.2 and Alibaba’s Qwen 3.8-Max in reasoning and coding tasks. Performance metrics include a score of 80.9 on SWE-Bench Verified and 80.1 on Terminal Bench v2.1. Reflection AI asserts that the model requires 3 to 4 times less inference compute than its rivals.
Reflection AI has secured over $4 billion in funding, with a recent pre-money valuation reaching $25 billion. Key investors include Nvidia, Sequoia, and Citigroup. The company maintains partnerships with the Pentagon and the US Department of Energy, focusing on enterprise-grade AI solutions.
The full model weights and documentation are scheduled for release in late October 2026 under the Apache 2.0 license. This launch aims to provide a Western alternative to Chinese open models, which currently account for over 30% of token share in the market.
