This repository contains an improved EAGLE3 draft model (Eagle3.1) designed for speculative decoding alongside the Kimi-K2.6-NVFP4 target model. It incorporates fc_norm and norm_output enhancements that improve benchmark performance across coding, math, and multilingual tasks. The model is intended for developers optimizing LLM inference speed.
Kimi K2.6 Eagle3.1 Mla is a Foundation models & chat project. It focuses on accelerating inference speed of large Kimi models through effective speculative decoding with a specialized draft model. It is built as an open-source project for developers. Kimi K2.6 Eagle3.1 Mla is open source under the Apache-2.0 license. It ships for the web, the command line, and API.
LightSeek Foundation builds and maintains Kimi K2.6 Eagle3.1 Mla, and it first shipped in 2023. Development happens publicly on GitHub with 87.9k stars and 3.2k commits in the last 90 days. Among its 3 catalogued features are speculative decoding, EAGLE3 architecture, and improved normalization.
Summary written by a language model from the project’s public pages.
What PulseGate has recorded for this listing
Closest matches by what these projects do