Beam launches 501B open-weight model

Beam is a significant Western open-weight bet, but its headline benchmark claims remain unverified until developers can access the weights and reproduce the results.

• –Its 23B active-parameter design aims to deliver frontier capability with substantially lower inference costs than larger rivals.
• –Reflection trained Beam on 23.8T tokens and more than 100 million RL rollouts, signaling unusually large investment in agentic training infrastructure.
• –Apache 2.0 licensing and planned integrations could make Beam attractive for enterprise customization and self-hosted deployments.
• –The model is currently a waitlisted preview, so hardware requirements, serving economics, safety behavior, and real-world coding performance remain open questions.
• –Early community reaction is positive about the license and Western provenance, while skepticism centers on the lack of independently verified evaluations. [TechCrunch](https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/) [LocalLLaMA discussion](https://www.reddit.com/r/LocalLLaMA/comments/1wyik2p/reflection_ai_announced_beam_501b_openweight_model/)