reflection/Beam-1.0-501B-A23B model card

Fetched 2026-10-18. Weights repo created 2026-10-07 (HF API createdAt 2026-10-07T19:52:01Z), last modified 2026-10-16. Card nameplate: sparse MoE, 501.3B total parameters, 23.4B active, 52 layers (3 dense + 49 MoE), gated GQA plus sliding-window attention, 1M context, vocabulary 152k tokens, 256 experts, Apache-2.0. The card's benchmark table is a vendor claim with BF16, FP8 and NVFP4 columns. BF16 column: SWE-bench Verified 81.6, SWE Bench Pro v1 66.1, SWE Bench Pro v2-Hard 76.5, SWEBench Multilingual 78.3, Terminal Bench v2.1 80.1, GPQA Diamond 90.6, AIME 2026 97.7, LiveCodeBench 88.3, HLE without tools 36.0. HF safetensors metadata reports 501,203,374,528 BF16 parameters and 64,237,824 F32 parameters. The card has no engine or serving section.