| license: mit | |
| # kimi-k2.5-eagle3 | |
| Post-trained EAGLE3 draft model for [moonshotai/Kimi-K2.5](https://huggingface.co/moonshotai/Kimi-K2.5), based on [lightseekorg/kimi-k2.5-eagle3](https://huggingface.co/lightseekorg/kimi-k2.5-eagle3). | |
| ## Overview | |
| This model is a speculative decoding draft model trained using the [EAGLE3](https://arxiv.org/abs/2503.01840) architecture. It can be used to accelerate inference of Kimi-K2.5 by predicting multiple tokens in parallel. | |
| ## Training | |
| The base EAGLE3 draft model from [lightseekorg/kimi-k2.5-eagle3](https://huggingface.co/lightseekorg/kimi-k2.5-eagle3) was further post-trained on open-source coding datasets. | |
| ## Usage | |
| This model is intended to be used as a draft model with EAGLE3-compatible inference engines such as [vLLM](https://github.com/vllm-project/vllm) or [SGLang](https://github.com/sgl-project/sglang). | |