deepseek-v4.1IMAGE

Adaptive Experts boosts DeepSeek V4.1 Flash inference on GB300

The author applied an Adaptive Experts technique to a vLLM deployment of DeepSeek Flash 1017 on a DGX Station GB300, reporting a solid speedup without a sidecar and expecting it to transfer to other MoE models, with a GitHub recipe linked.

James Meadlock

@JamesMeadlockArchived 2 days agoOriginally posted 2026.10.27

Adaptive Experts boosts DeepSeek V4.1 Flash inference on GB300

James Meadlock @JamesMeadlockOriginally posted 2026.10.27

Open original on X

Adaptive Experts! Nice performance gain for GB300 (no sidecar) for Deepseek Flash 1017; technique should work on other MoE models. https://github.com/J-M-Recipes/recipes/tree/main/recipes/dgx-station-gb300/deepseek-v4.1-flash-vllm-uva-dspark

x.com/jamesmeadlock/status/2114850707233022086Archived 2026.10.27 12:08