Reflection AI drops Beam, a 501B coding powerhouse

Beam is a massive new open-weight model from Reflection AI, designed specifically to tackle software engineering, multi-step reasoning, and agentic workflows.

It is built on a sparse Mixture-of-Experts (MoE) architecture with 501B total parameters, but it only activates 23B parameters per token. It features a 1 million token context window and was pretrained on 23.8 trillion curated tokens. Basically, it gives you the knowledge capacity of a half-trillion-parameter beast while keeping per-token compute down to the level of a much smaller model.

How they pulled it off

Beam’s muscle comes from an unprecedented investment in high-compute reinforcement learning (RL).

Reflection threw 10,500 Nvidia GB300 GPUs at a four-week continuous training run. That fleet generated over 100 million rollouts across 1.3 billion sandboxes, tapping into a pool of roughly one million specialized environments spanning competitive coding, command-line operations, and STEM problem-solving.

Running RL at that scale usually breaks down due to policy staleness, where models try to learn from data generated by older checkpoints. Reflection engineered an asynchronous policy gradient system that stays numerically stable even when training on trajectories generated over 24 hours earlier. They also built a custom hierarchical weight distribution pipeline that pushed new checkpoints to the entire inference fleet in a median of 12 seconds, slashing cross-rack traffic by 75%.

Under the hood, Beam balances its experts using cosine decay on bias updates without auxiliary losses, ensuring uniform expert utilization so no compute capacity sits idle.

What it can do

On coding and advanced reasoning evaluations, Reflection claims Beam scores on par with Z.ai’s GLM 5.2 and approaches heavyweight models like Qwen 3.8-Max. More importantly, it hits those GLM 5.2-level scores while using 3 to 4 times less inference compute.

Although Beam is text-only, its reasoning generalizes across domains. Without explicit browsing data in its RL mix, the model organically learned to query web tools, chain external APIs, and trigger OCR endpoints to read documents when sandboxed.

Developers also get a native reasoning effort toggle. You can dial it down for concise, rapid responses to save budget, or crank it up to let Beam deliberate through complex debugging and architectural tasks.

How to get it

Beam is currently undergoing final red-teaming. Early access signups are open now on Reflection's website, and the team plans to release the model weights, technical report, model card, and developer artifacts later this month across major open-source hubs and cloud platforms.

Why it matters

If you build autonomous coding tools or run enterprise developer platforms, Beam is one to watch. A 501B footprint means local deployment still requires serious hardware, but activating just 23B parameters turns it into a genuine workhorse that delivers frontier coding intelligence at a fraction of typical MoE serving costs.

We think Beam’s real achievement is validating brute-force, high-compute RL as an efficiency play, not just a brute-intelligence benchmark chaser. If the released weights deliver on the promised 3 to 4 times compute savings, it could reset expectations for what high-throughput open coding agents look like.