Original sourceReflectionAdditional: the-decoder
Summary
Reflection released its first open-weight model, Beam, using a sparse mixture-of-experts (MoE) architecture with 501B total parameters and only 23B activated per token, targeting coding, reasoning and agentic workloads. The company says Beam performs comparably to GLM 5.2 on hard reasoning tasks wh…
Key points
- This is one of the few large open-weight models released under Apache 2.0 outside China, offering a real alternative for development teams that need to self-host, fine-tune or control costs.
- Beam emphasizes compute efficiency among open-weight models with its sparse architecture, potentially changing the cost structure for enterprises building their own agents and coding assistants.
- Development teams can deploy and fine-tune it themselves after weights are released this month, reducing reliance on closed APIs and using existing open-source toolchains for evaluation and integration.
Editorial note
This page is Code & Chain's editorial summary of public sources. It may be prepared with AI assistance and published through an automated workflow. Refer to the original sources; this content is not investment, legal, or tax advice.