What they found
The authors describe Beam, a sparse Mixture-of-Experts model. That means only part of the model runs for each input: 23.4B of 501.3B total parameters. They say it was built for coding, reasoning and agentic work, where an SI system carries out multi-step tasks. They report that it reaches frontier level reasoning while needing 3-4x less inference compute than many models of its class. It was pretrained on 23.8T tokens and has a 1M-token context window, made possible by a two-stage midtraining phase. Post-training used a fully asynchronous reinforcement learning run on 10.5K GB300 GPUs. It produced over 100M rollouts over 4 weeks. The authors say capabilities kept improving as RL compute grew, with no sign of a plateau. On safety, they report that Beam refuses or pivots on 88.9% of harmful requests and complies with 81.5% of harmless ones.
Why it matters
If the claims hold, strong reasoning may be served at a much lower cost per answer. That makes it easier to deploy SI systems that work on long coding and agent tasks. The report also says RL gains had not levelled off at this scale. That suggests more RL compute may keep paying off. The abstract gives no benchmark names or scores, so the strength of these claims cannot be judged from it.
How it changes the race
The authors say an open lab ran one of the largest RL campaigns to date and reached frontier level reasoning at lower serving cost. If that holds, the gap between open and closed frontier work narrows. It also adds evidence that scaling RL compute is a path to better capability.
Safety and control
The authors report that Beam has the highest refusal or pivot rate on harmful requests among open-weight models on the Pareto frontier. That is, it balances refusing harm against helping with harmless requests. The abstract does not say how this was tested, so the result should be read with care.
The abstract, by the authors
Beam is a sparse Mixture-of-Experts (MoE) model with 501.3B total parameters and 23.4B active parameters, specifically built and optimized for coding, reasoning, and agentic workloads. Beam delivers frontier level reasoning, while requiring 3-4x less inference compute than many models of its class. Pretrained on 23.8T diverse, curated, high-quality tokens, Beam features a 1M-token context window achieved through a novel two-stage midtraining phase, built on verifiable, correctness-anchored data. Post-training leveraged a large-scale, fully asynchronous reinforcement learning (RL) campaign across 10.5K GB300 GPUs, generating over 100M rollouts over 4 weeks of training. We believe this is one of the largest scale RL runs conducted by any open lab to date. Across our evaluation suite, capabilities continued to improve as we increased RL compute, with no sign of a plateau. Supported by deep hardware-model co-design, FP8/NVFP4 quantization, and a 1.93B-parameter speculative decoding drafter (DFlash 2), Beam maximizes serving throughput with efficient, flexible deployment. Averaged across conversation turns and harm categories, Beam refuses or pivots on 88.9% of harmful requests while complying with 81.5% of harmless requests, achieving the highest refusal/pivot rate on harmful requests among open-weight models on the Pareto frontier.
Quoted from arXiv:2611.26358. The research and its words belong to its authors.
