first open-weight model

Beam, Reflection's first open-weight model, is introduced. This sparse Mixture-of-Experts model boasts 501 billion total parameters, with 23 billion active, designed for coding, reasoning, and agentic workloads. Beam's performance stems from substantial investments in both pretraining and reinforcement learning (RL). Pretraining involved curating 23.8 trillion high-quality tokens from the web and proprietary licensed datasets, matching or surpassing similar-sized open base models.

Simultaneously, RL algorithms, training environments, and infrastructure were developed to sustain high-compute RL at an exceptional scale. By generating over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks, Beam achieved competitive open-weight performance with frontier inference compute efficiency. Beam is currently undergoing final red-teaming and evaluations.

Early access to the model can be signed up for. The model focuses on coding and agentic performance, surpassing larger open models like GLM 5.2 and approaching Qwen 3.8-Max in coding and agentic tasks. Beam's efficiency at inference time sets it apart from frontier open models like Kimi K3. The model demonstrates high-efficiency reasoning, achieving scores comparable to GLM-5.2 while using 3-4x less inference compute.

Moreover, Beam's performance improves as RL compute increases, with no signs of plateau. Asynchronous policy gradients were used to train Beam, addressing stability challenges arising from long-running rollouts and numerical mismatches between training and inference engines. Beam's RL training also enables controllable length penalties, allowing users to balance capability and token usage according to their task and compute budget.

The model exhibits the ability to generalize beyond its training tasks, learning broad agentic capabilities through training on reasoning, software engineering, and terminal tasks. Beam's capability to work with information from other modalities when represented as text is demonstrated in various out-of-distribution domains, such as building a live NYC subway dashboard, creating interactive applications, and preparing model fine-tuning notebooks.