Beam: Efficient reasoning for coding and agentic workloads

Baseten has introduced day-one support for Beam, Reflection’s first open-weight model, through its Model APIs and Dedicated Inference offerings. Beam is a 501-billion-parameter mixture-of-experts autoregressive transformer with 23 billion active parameters per token and a context window of up to one million tokens, designed primarily for English text generation, coding, reasoning, and agentic workloads. Reflection reports that the model performs strongly on coding, STEM, reasoning, and agentic benchmarks, with future models in the family expected to add multilingual and multimodal capabilities. Beam was trained extensively with reinforcement learning and long rollouts to handle multistep tasks and tool use, while a length penalty encourages efficient solutions that use fewer tokens. Users can tune its reasoning-effort setting to prioritize faster, shorter outputs or deeper reasoning for more difficult tasks, and the model is intended to work with established developer runtimes, frameworks, coding harnesses, and tools.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.