StorySoftware
Reflection AI announces Beam, 501B-parameter open-weight model for coding and reasoning
Reflection AI announced Beam, a 501-billion-parameter sparse mixture-of-experts open-weight model for coding, reasoning and agentic tasks. With only 23 billion parameters active per token, Beam achieves 3–4x more efficient inference than comparably capable dense models, with performance matching GLM-5.2 on advanced reasoning. Weights, technical report and developer tools will release under Apache 2.0 in October 2026.
- Beam: 501B total parameters with 23B active per token via sparse mixture-of-experts design
- Achieves 3–4x more efficient inference than larger dense models while matching GLM-5.2 on reasoning benchmarks
- Founded March 2024 by Misha Laskin and Ioannis Antonoglou, former DeepMind researchers on Gemini and AlphaGo
- Raised $2B in Series B at $8B post-money valuation (October 2025), led by NVIDIA and Lightspeed; $130M in seed/Series A in March 2025
- Weights and technical documentation to release under Apache 2.0 in October 2026; trained on 23.8T tokens plus 100M+ RL rollouts on 10.5K NVIDIA GB300 GPUs