A community-published variant of Kimi K2.5 that incorporates Eagle3 speculative decoding and Multi-head Latent Attention (MLA), two architectural optimizations aimed at faster inference throughput. Published by lightseekorg in safetensors format, it targets researchers and developers experimenting with efficient serving of Kimi-family models. As a derivative focused on inference speed rather than capability improvements, expect similar quality to the base Kimi K2.5 with potentially lower latency on compatible serving frameworks.