DSpark speculative decoding for Kimi K3.1 Pro cuts draft KV memory
AI-generated summary Inferact and Red Hat's AI Speculators team released DSpark, a speculative decoding draft model for Moonshot and Kimi's Kimi K3.1 Pro. The new model utilizes a 2K sliding window in every draft layer to bound draft KV memory when operating at one-million-token contexts.