Reflection open-sources 501B MoE Beam; vLLM day-0 support hits 708 tokens/s
Xianbao_QIAN · x · 2026-10-16
Reflection AI released Beam-1.0-501B-A23B, an open-source 501B-parameter MoE with 23B active parameters per token, supported by vLLM from day 0 and optimized on NVIDIA Blackwell.
Key points
- Architecture: 52-layer MoE decoder (hidden size 5120) with DeepSeek-V3-style routing; hybrid attention mixing full attention with NoPE (SSMax) at select layers and sliding-window (512) elsewhere, GQA with 48 query / 8 KV heads, plus sandwich post-norm and output gating (see architecture diagram).
- Agent-driven kernel optimization: coding agents autonomously tuned the kernels, making batch-1 decode 2.71× faster and prefill 1.87× faster than the first implementation.
- Speculative decoding: the DFlash2 drafter reaches 708 tokens/s on a single request at TP8 with 3.7 acceptance length; adaptive verification adds 10% throughput at high concurrency.
- Out of the box: reasoning, tool calling, and structured outputs work by default; tensor, data, and expert parallelism supported; NVFP4 quantized release.
- Deployable via the vllm/vllm-openai:reflection-beam Docker image.
Related event: Reflection AI open-sources 501B MoE model Beam with 40+ partner ecosystem(20 posts)→
More from Infra
- Adobe ships NVIDIA RTX Spark optimizations: up to 2x faster creative apps and 5x Frame.io gains — shashib · 2026-10-16
- Meta's commitments jumped to $349B in a quarter as server memory prices tripled ahead of Oct 21 earnings — tengyanAI · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- Meta's open Muse Spark needs 493GB of memory — far beyond any consumer laptop — TejasKumar_ · 2026-10-16
- CoreWeave plunges 20% as credit markets price >60% default odds; is this dot-com all over again? — kimmonismus · 2026-10-16
- MiniCPM-V 4.7 35B-A3B hits 17+ tokens/s on iPad with ~3GB memory — techNmak · 2026-10-16