SYLIN

Kimi K3 DSpark

by Inferact · Model · Language & code models

4B draft model for speculative decoding of Kimi K3, drafting seven tokens per pass, on Hugging Face

AccessOpen weightsSize4B parametersLicensekimi-k3

Evidence

  1. Jul 29
    Reaches 464 tokens/s versus roughly 5-15 tok/s for other local methods; needs four GB300 GPUs minimum

Alternatives

Made from