vLLM adds day-zero Reflection Beam serving after agent-driven kernel optimization

vLLM has added day-zero serving support for Reflection AI’s Beam-1.0 model, pairing standard serving features with a set of model-specific kernel and speculative-decoding optimizations. The vLLM engineering post says Beam is a 501-billion-parameter mixture-of-experts model that activates 23 billion parameters per token and exposes a one-million-token context window.

What the implementation supports

The initial integration supports reasoning output, tool calling and structured outputs, with tensor, data and expert parallelism available as deployment choices. vLLM also published a container image and launch configuration for Beam’s NVFP4 checkpoint, including FP8 KV cache and the model’s reasoning and tool-call parsers. Those details are documented in the project’s day-zero serving guide.

Beam’s architecture mixes 43 sliding-window attention layers with nine full-context layers that omit positional encoding. The model uses 256 routed experts, selecting eight per token, while its full-context KV cache is limited to those nine layers. According to vLLM’s architecture breakdown, that design keeps a one-million-token FP8 KV cache at roughly 18 GiB before sharding.

Agents worked on the serving path

The notable part of the release is how the first implementation was optimized. vLLM says coding agents profiled the model, wrote kernels and benchmarked each change. On an eight-way tensor-parallel GB200 setup, the project reports that seven changes reduced batch-one decode time from 8.95 milliseconds to 3.30 milliseconds per step and cut an eight-sequence, 2,048-token prefill from 328 milliseconds to 176 milliseconds. The published optimization sequence includes fused normalization and residual kernels, fused communication, a specialized FP32 router kernel, merged projections and a fused KV-cache update.

Those are project-reported measurements, not an independent comparison. They nevertheless make this more than a compatibility announcement: the post identifies the operations changed, the measured configuration and the latency movement after each stage.

Speculative decoding under load

Beam ships with a DFlash2 draft head, which vLLM uses to propose a block of tokens for verification in one pass. The implementation can adapt the number of draft tokens it verifies based on estimated acceptance and cost. In the project’s SPEED-Bench test, adaptive verification was essentially flat at concurrency one and eight, then improved output throughput by 6.9% at concurrency 32 and 10.8% at concurrency 128. vLLM cautions that the test used a low-entropy subset dominated by code completion and text sorting, so open-ended chat may have lower acceptance.

For platform teams, the practical takeaway is that Beam can be evaluated through vLLM’s existing OpenAI-compatible serving path, but the strongest performance numbers depend on specific Blackwell hardware, parallelism choices and model-specific kernels. The published container makes the path reproducible; production sizing still needs workload-specific testing.