vLLM Ships Day-0 Support and Optimizations for Reflection Beam
On October 16, the vLLM team rolled out Day-0 support for the Reflection Beam model along with a suite of accompanying optimizations. Beam is a sparse MoE model with 501B total parameters that activates only 23B per token; vLLM provides the Docker image vllm/vllm-openai:reflection-beam, an official recipe, and a blog post. In a chat benchmark at concurrency 16, a 4×B300 NVFP4 TP4 configuration achieved 1,769 tok/s output.
Confirmed
- Architecture details: Beam 1.0 has 52 layers, with the first 3 using dense MLPs and the rest MoE; of 256 experts, 8 are activated per token. Attention alternates between a 512-token sliding window and global NoPE full attention, balancing local processing with long-context access
- Day-0 RL support: vLLM partnered with SkyRL, covering LoRA, BF16 full fine-tuning, and MXFP8 RL, with vLLM rollout and Reflection's Megatron-Bridge implementation; by adapting expert placement in vLLM, LoRA decoding throughput improved by 62%
- DFlash2 decoding acceleration: Reflection reports that DFlash2 delivers 1.72–3.03× pure decoding speedups across GSM8K, MATH500, HumanEval, MBPP, and MT-Bench at concurrency 1/16/32, with the highest speedup of 3.03× on GSM8K at concurrency 1
- Long-context kernel optimization: Inferact performed kernel-level optimizations for Beam's long-context serving, including fused all-reduce + norm, a packed QKV path, and tanh softcap kernels; in GB200 forward microbenchmarks, TP4 batch 1 saw speedups of up to about 20%
Why it matters
As an ultra-large sparse MoE model, Beam's accessibility to the open-source community hinges on out-of-the-box support for its inference and training stacks. vLLM's full-pipeline optimization—from serving deployment (Docker image and recipe) to reinforcement learning (SkyRL Day-0 support) to low-level kernels (fused kernels, DFlash2)—demonstrates the complete engineering scaffolding needed to put large MoE models into production, and offers a reference blueprint for deploying similar models.
2026-10-16 ~ 2026-10-16 · 10 related posts
- Episode 1: Reflection AI Open-Sources Beam 1.0, a 501B-Parameter MoE Model(2026-10-16, 24 posts)
- Episode 2: vLLM Ships Day-0 Support and Optimizations for Reflection Beam(2026-10-16, 10 posts)
Primary sources
- Inside Beam: 8-of-256 expert routing across 52 layers, sliding window plus NoPE attention — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- [source] DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- vLLM serves Beam at 1,769 tok/s on 4x B300 NVFP4, publishes Docker image and recipes — vllm_project · 2026-10-16
- Fused kernels cut GB200 forward latency up to 20% in Beam serving microbenchmarks — vllm_project · 2026-10-16
- [source] vLLM day-0 support for Reflection Beam: 1,769 tok/s on 4x B300 NVFP4 — vllm_project · 2026-10-16
- [source] vLLM ships Day-0 RL support for Beam in SkyRL, +62% LoRA decode throughput — vllm_project · 2026-10-16
2 near-duplicate retellings: vllm_project · vllm_project