Reflection opens Beam, its first open-weight model: 501B parameters, 23B active, fewer tokens per coding task
Synopsis
Reflection released Beam, its first open-weight model, with day-0 support on the Baseten platform. Beam is an autoregressive transformer with a mixture-of-experts architecture: 501B total parameters, 23B active per token, and a context window up to 1 million tokens, aimed at coding and agentic workloads. Training leaned on reinforcement learning, using longer multi-step rollouts so the model can practice tool use and adjust from feedback, plus a length penalty against unnecessary tokens, with the goal of cutting the cost of these workloads.
Interpretation
Reflection introduced Beam, the company's first open-weight model, and brought it to the Baseten platform with day-0 support, so developers can call it through Baseten Model APIs and Dedicated Inference. Reflection previously had no open-weight model available for outside use; Beam is the first member of the family, with multilingual and multimodal capabilities listed as plans for future releases. The model is available now through Baseten Model APIs and Dedicated Inference, and the announcement states it works out of the box with the runtimes, frameworks, coding harnesses, and tools developers already use.
Beam was trained to solve coding and agentic tasks while using fewer tokens to reach a successful result, which lowers the cost of those workloads. The usual target is simply solving the task; Beam also optimizes how many tokens it spends getting to a successful outcome. Reflection's evaluations cover coding, agentic, reasoning, and STEM benchmarks and are described as strong; during training, successful solutions were rewarded and a length penalty discouraged unnecessary tokens.
Beam's reasoning effort is adjustable, with lower settings favoring shorter responses and higher settings allowing more reasoning to improve performance on demanding tasks. This lets one model trade token savings against performance on hard tasks instead of fixing a single reasoning length. The adjustment is exposed as a reasoning effort parameter, and the model is described as intended for developers and researchers building coding and agentic systems, as well as enterprises that want to customize and run their own models.
Perspective
The model targets developers and researchers building coding and agentic systems, plus enterprises that want to customize and run their own models; it is available through Baseten Model APIs and Dedicated Inference and works with existing runtimes, frameworks, coding harnesses, and tools. Multilingual and multimodal capabilities are stated as plans for future releases.
The performance description comes from Reflection's own evaluations, and the material lists no specific benchmark names, numbers, or comparison models; the material also gives no quantified result for how much the length penalty and reasoning effort parameter actually save in tokens.