Reflection Releases Open-Weight Beam Model, Cuts Compute Costs

AI startup Reflection released an open-weight model called Beam. The company was founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Beam has 501 billion parameters, but only 23 billion of them activate per token. Reflection says Beam matches the Chinese model GLM 5.2 on key benchmarks. The company also says it uses three to four times less compute. The weights will ship under the open Apache 2.0 license. A technical report will follow later this month.

Cutting compute three to four times compared with GLM 5.2 is a real efficiency gain. The race for efficiency is no longer limited to Chinese labs. But Reflection's own benchmarks show that stronger open-weight models, including Kimi K3, still beat Beam on raw performance. So Beam looks more like a step toward accessibility for teams without heavy cloud infrastructure. It is not a reason to switch away from Opus or Fable, Anthropic's AI models, for everyday agentic coding work.