Reflection AI Beam 501B: Can Four B200s Carry 1TB of Weights?
Reflection AI's first open-weight model Beam actually landed on 2026-10-16: repository huggingface.co/reflection/Beam-1.0-501B-A23B (plus FP8 and NVFP4 variants) under Apache 2.0, with the model card marked informational only and non-binding. The timeline matters: the 10-05 announcement came with only a waitlist and a promise of weights this month, a benchmark table update on 10-08, no public repository on the HF org page as of 10-13, then weights, Hugging Face's own blog post and an AWS SageMaker path all landing on 10-16 - eleven days from promise to delivery. Architecture: a sparse MoE with 501B total parameters, 23B active per token (about 4.6 percent), 52 layers, 256 experts with 8 routed per token, text in and text out; the 1M effective context stretched during midtraining and the 256K RL rollout length are two different figures that must not be merged, and the serving-side context ceiling and max output are unpublished. Reasoning depth has five levels (low to max) and the weights bundle a DFlash2 speculative-decoding drafter. The hardware bill is the core: BF16 at about 1TB, NVFP4 at about 280GB with Hugging Face saying two B200s fit barely and recommending four B200s (720GB); the smallest GGUF is about 103GB and is a RAM requirement first. All benchmarks are Reflection self-reported with zero independent replications: SWE-bench Verified 81.6 percent (FP8 80.2, NVFP4 80.7), Terminal-Bench v2.1 80.1, GPQA Diamond 90.6, AIME 2026 97.7, HLE without tools 36.0; Artificial Analysis has indexed it at Intelligence Index 30, second among US open-weight models, relayed secondhand. Ollama offers day-zero free access but ollama run beam pulls beam:cloud, its hosted inference - local runs need an explicit local GGUF. Mirror CLI (Rust) ships as an open agent harness alongside 40-plus partner integrations; the self-distillation recipes remain unpublished, and the FT report that Nvidia is in talks to acquire or deepen its stake in Reflection is unconfirmed by both parties.