vLLM trains a DSpark speculator for 2.8T-param Kimi K3, hitting ~435 tok/s

AccBalanced · x · 2026-09-18

The vLLM team trained a DSpark speculator for Kimi K3, a 2.8T-parameter model, using the DFlash block-level speculative decoding extension from DeepSeek.

Results: single-stream math reasoning jumps from 110 to 435 tok/s/user, and up to 3.5× higher output throughput at matched interactivity under concurrent load. The 5B drafter proposes 8 tokens per step with 6.4 accepted on math.

Training at this scale required multi-node setups: the Speculators library plus a new Mooncake connector streaming hidden states between vLLM inference and training over RDMA—two GB300 NVL72 nodes serving the target, one training the drafter. The recipe is validated on Qwen3.6-35B-A3B, Gemma-4-31B-it, GLM-5.2, and ships in a HF-compatible format vLLM loads directly.

Original post →

More from Infra

Infra channel →