Beam: Reflection's 501B open-weight model
Article URL: https://reflection.ai/blog/introducing-beam Comments URL: https://news.ycombinator.com/item?id=49969183 Points: 171 # Comments: 50
Beam is Reflection's first open-weight model, boasting 501 billion total parameters with 23 billion active. This sparse Mixture-of-Experts model is specifically designed for coding, reasoning, and agentic workloads.
The model's development involved substantial investments in both pretraining and reinforcement learning (RL). Pretraining occurred on an extensive 23.8 trillion tokens sourced from the web and proprietary licensed datasets, surpassing similar-sized open base models. Concurrently, the algorithms, training environments, and infrastructure necessary for high-compute RL were developed.
The RL phase of training involved over 100 million rollouts, executed on 10.5K NVIDIA GB300 GPUs over four weeks. This methodology yielded open-weight performance with exceptional inference compute efficiency. Beam is currently undergoing final red-teaming and evaluations, with the technical report, model card, and developer artifacts slated for release later this month.
Beam's primary focus during training was on coding and agentic performance, positioning it competitively with larger open models like GLM 5.2 and Qwen 3.8-Max, particularly in terms of efficiency at inference time. While frontier open models like Kimi K3 lead in raw capability, Beam's advantage lies in its efficiency during inference.
Beam demonstrates strong coding and agentic capabilities coupled with highly efficient reasoning. On advanced reasoning benchmarks, it achieves scores comparable to GLM-5.2 while using 3-4 times less inference compute. This efficiency advantage is even more pronounced when compared to models in the 2T+ parameter family, such as Qwen 3.8-Max, which require significantly more inference compute per token.
These efficiency gains translate into delivering stronger model capabilities at a lower cost, making Beam a powerful workhorse model for enterprise coding and agentic workloads. The RL training approach involved asynchronous policy gradients, addressing the challenge of numerical mismatch between training and inference engines. Through new algorithms, Beam ensures stable learning even when learning from interactions generated more than a day earlier.
The training process also incorporated a controllable length penalty, encouraging successful solutions while discouraging unnecessary tokens. This allowed the model to learn more effectively with less reasoning initially, transitioning to stronger agentic capabilities as it developed. Beam's RL training enabled it to generalize beyond its training tasks, demonstrating unexpected abilities like browsing, tool use, and even creating fine-tuning notebooks for other models.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- Beam: Reflection's 501B open-weight model reflection.ai