Reflection previewed Beam, a 501B model it says it will open-weight under Apache 2.0
Reflection introduced Beam on Monday: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads. It was pretrained on 23.8 trillion tokens and holds a 256K-token context. The company says Beam is competitive with GLM-5.2 and approaching Qwen 3.8-Max on coding and agentic benchmarks while using three to four times less inference compute, and that it will release the weights under Apache 2.0 this month.
The scale of the training run is part of the claim. Reflection says its reinforcement-learning stage generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks, against a pool of nearly a million environments and about 1.3 billion sandboxes - which it calls among the largest RL runs by any open lab. Beam is text-only and not downloadable yet: it is still in red-teaming, the waitlist is open, and the technical report is promised alongside the weights.
WHY IT MATTERS
Open-weight models have been closing on the frontier from outside the United States for two years; this is the most serious American attempt yet to answer them, and the efficiency pitch is the part that matters commercially - near-GLM-5.2 reasoning at a fraction of the tokens is a self-hostable workhorse rather than a demo. It also reframes the argument in Washington, where open weights have been filed as a Chinese lead to be closed rather than a distribution choice to be made. Reflection is candid that Kimi K3 stays ahead on raw capability, and the caveat deserves one line: the comparison is Reflection’s own and nothing is downloadable yet.