Kimi K3.1 Pro DSpark model card: 2.5B draft, 5.57 acceptance length on GSM8K

vllm_project · x · 2026-11-18

vLLM published weights and ready-to-use configs for Inferact/Kimi-K3.1-Pro-DSpark: a 2.5B, 5-layer draft model with 2048-token sliding-window GQA per layer, drafting 7 tokens per forward pass and sharing embeddings with the target model. Benchmarks (k=7, temp 1.0, topp 0.95) show acceptance lengths of 5.566 on GSM8K, 4.660 on HumanEval, 4.376 on MBPP, with bounded draft KV cache up to 1M-token context.

Original post →

More from Infra

Infra channel →