Reflection's Beam open-weight model launches on Baseten with day-0 support
Baseten has added Beam, the first open-weight model from Reflection, to its platform with day-0 support. Beam is a mixture-of-experts model aimed at coding and agentic workloads, and it is trained to reach results with fewer tokens.
- Baseten supports Reflection's first open-weight model, Beam, from day 0.
- Beam is an MoE model with 501B total parameters, 23B active per token and a context window of up to 1 million tokens.
- Reflection trained it with RL and a length penalty to solve coding and agentic tasks with fewer tokens.
- A reasoning effort parameter lets users trade shorter responses for more reasoning.
- Available via Baseten Model APIs and Dedicated Inference.
What Beam is and how it is available
According to Baseten, Beam is now available through Baseten Model APIs and Dedicated Inference. The company said it is proud to support Reflection by bringing the model to its platform on the day of release.
Beam is an autoregressive transformer that accepts text inputs and generates text outputs, mainly in English. It uses a mixture-of-experts (MoE) architecture with 501B total parameters and 23B active parameters per token. Its context window reaches up to 1 million tokens.
Beam is the first in a new family of models from Reflection. According to Baseten, multilingual and multimodal capabilities are planned for future releases.
Built for efficient reasoning
Reflection trained Beam to solve coding and agentic tasks while using fewer tokens to reach a successful result. The stated goal is to reduce the cost of these workloads. Reflection's own evaluations, as cited by Baseten, show strong performance across coding, agentic, reasoning, and STEM benchmarks. The source does not include the benchmark figures in text form.
Reinforcement learning (RL) was a major part of the training. Reflection used longer RL rollouts so that Beam could practice solving tasks over multiple steps, using tools and adjusting its approach based on feedback. Successful solutions were rewarded, and a length penalty discouraged unnecessary tokens.
- Longer rollouts let the model work through complex tasks.
- The length penalty encourages it to do so with fewer tokens.
Users can also adjust a reasoning effort parameter. Lower settings favor shorter responses, while higher settings allow more reasoning to improve performance on demanding tasks.
Who it is for
Baseten says Beam is designed for developers and researchers building coding and agentic systems, as well as enterprises that want to customize and run their own models. The model works out of the box with the runtimes, frameworks, coding harnesses, and tools developers already use, according to the company.
For teams running agent-style workloads, where long multi-step tasks can consume many tokens, the combination of an open-weight release, a configurable reasoning effort and a training approach that penalizes wasted tokens is aimed at keeping costs down. Baseten's pitch is that users can build coding and agentic systems that solve complex tasks without wasting tokens. Pricing and performance numbers were not given in the announcement.
Baseten: Beam: Efficient reasoning for coding and agentic workloads
