Beam: Reflection's 501B open-weight model

Beam activates 23 billion of its 501 billion parameters per token, so roughly one in 22 does work on any given token. Reflection pretrained it on 23.8 trillion tokens in under four weeks on Nvidia GB300 NVL72 systems, then ran reinforcement learning for four weeks on about 10,500 GPUs. The company says the context window reaches one million tokens. Reflection claims it matches GLM 5.2 on reasoning with 3 to 4 times less inference compute. The benchmark picture is mixed. Kimi K3 remains ahead on raw capability, and Reflection pitches Beam on inference efficiency. DeepSeek V4.1 Flash posts the top scores on some individual tests, including Terminal-Bench v2.1 and DeepSWE. The target is a Western alternative in a field dominated for a while by Chinese labs such as GLM, Kimi, Qwen and DeepSeek. Beam is in final red-teaming, with early access through a waitlist. The API is OpenAI-compatible, and no price has been published. For founders, the open question is cost. A permissive license on a US-built model could suit buyers who avoid Chinese-origin weights, and the sparse design points to lower serving bills. Both claims are vendor-reported. Self-hosting, quantizing and fine-tuning wait for the weight release, so independent evaluations and real per-token pricing should come before any migration decision.