Reflection’s Beam: an open model you can run on your own servers with lower compute needs

Reflection has introduced Beam, its first open-weight model, and the pitch is straightforward: big model, smaller appetite. Beam is a sparse mixture-of-experts system with 501 billion total parameters but only 23 billion active at a time, aimed at coding, reasoning, and agentic work.

The company says that tradeoff came from brute-force pretraining and an unusually heavy reinforcement-learning push. Beam was pretrained on 23.8 trillion curated tokens from the web and licensed proprietary data, then pushed through more than 100 million rollouts across 10.5K NVIDIA GB300 GPUs over four weeks. Reflection says that training run used roughly 1.3 billion sandboxes and that capabilities kept improving as RL compute scaled up, with no sign of a plateau.

The interesting part is not just that Beam can do the work. It’s that Reflection is arguing it can do it with less inference compute than comparable open models. The company says Beam is competitive with models such as GLM 5.2 on coding and agentic tasks, approaches Qwen 3.8-Max in those areas, and matches GLM-5.2 on advanced reasoning while using 3–4× less inference compute. It also says Beam is less capable than Kimi K3 on raw capability, but more efficient at serving time.

That efficiency story extends into the training stack. Reflection built an asynchronous RL system with stable learning under policy staleness, fast weight updates, and enough infrastructure to keep an average of 110K rollouts running at once. It says new weights reached the inference fleet in about 12 seconds median, and that 71 inference incidents were handled without stopping training.

Beam is still in final red-teaming and evaluation, with release of the weights, technical report, model card, and developer artifacts promised later this month. Early access is open now. For teams that care less about bragging rights and more about cost per useful token, that’s the real hook.