vLLM releases DSpark draft model for Kimi K3.1 Pro speculative decoding

vLLM, Inferact, and Red Hat AI released DSpark, a 2.5B draft model for speculative decoding of Kimi K3.1 Pro, achieving GSM8K acceptance length of 5.57 and up to 30% longer acceptance with bounded draft KV cache.

2026-11-18 ~ 2026-11-19 · 3 related posts

1 near-duplicate retellings: AccBalanced