Urgent.News

What's breaking now, across thousands of outlets.

AI

Beam: Reflection's 501B open-weight model

Beam, Reflection's inaugural open-weight model, has been unveiled. Beam is a sparse Mixture-of-Experts model boasting an impressive 501 billion total parameters, with 23 billion of those parameters active. This model is specifically designed for coding, reasoning, and agentic tasks.

The development of Beam involved significant investments in both pretraining and reinforcement learning (RL). The model was pretrained on 23.8 trillion diverse, curated, high-quality tokens sourced from the web and proprietary licensed datasets. This extensive pretraining enabled Beam to match or even surpass similar-sized open base models.

In parallel, the team developed advanced algorithms, training environments, and infrastructure to support high-compute RL at unprecedented scales. This effort involved over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over a period of four weeks. The combination of these efforts resulted in competitive open-weight performance with remarkable inference compute efficiency.

Beam is currently undergoing final red-teaming and evaluations. Interested users can sign up for early access to the model. Later this month, the weights, technical report, model card, and developer artifacts will be released.

During its training, Beam was focused on enhancing its coding and agentic performance. The model outperforms larger open models like GLM 5.2 and approaches Qwen 3.8-Max on coding and agentic tasks. However, where frontier open models like Kimi K3 excel in raw capability, Beam's strength lies in its efficiency during inference time.

Beam's performance across various coding, agentic, reasoning, and STEM benchmarks is showcased in the accompanying figure. This figure demonstrates Beam's ability to deliver strong model capabilities at lower costs, making it an ideal workhorse model for enterprise coding and agentic workloads.

Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Other reports of this story

Every outlet that reported this, and when →

Read the original at reflection.ai →

More in AI

More from Monday 5 October →