nvidia/Kimi-K2.6-Eagle3

Model Overview

Description:

The NVIDIA Kimi-K2.6 Eagle model is the Eagle head of Moonshot AI's Kimi-K2.6 model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA Kimi-K2.6 Eagle3 model incorporates Eagle speculative decoding with Model Optimizer.

With draft length 3, this Eagle head achieves an average acceptance length of 2.62 tokens per step on MT-Bench and 2.67 on the SPEED-Bench qualitative subset.

This model is ready for commercial/non-commercial use.

License/Terms of Use:

Governing Terms: Use of this model is governed by the NVIDIA Open Model License.

ADDITIONAL INFORMATION : Modified MIT License. Kimi-K2.6 .

Deployment Geography:

Global

Use Case:

Developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks where latency-optimized inference via speculative decoding is desirable.

Release Date:

Hugging Face 06/03/2026 via https://huggingface.co/nvidia/Kimi-K2.6-Eagle3

Reference(s):

• Kimi-K2.6 release notes
• EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test

Model Architecture:

Architecture Type: Transformers

Network Architecture: DeepSeek V3

Number of model parameters 1.8*10^9

Input:

Input Type(s): Text, Image, Video

Input Format(s): String, Binary(Base64 encoded), Binary(Base64 encoded)

Input Parameters: One-Dimensional (1D), Two-Dimensional (2D), Three-Dimensional (3D)

Other Properties Related to Input: Context length: 256k

Output:

Output Type(s): Text

Output Format: String

Output Parameters: One Dimensional(1D): Sequences

Other Properties Related to Output: Outputs may include natural-language responses, code, structured JSON, tool-call requests, agent coordination instructions, and generated artifacts depending on serving configuration and application-level tooling.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Supported Runtime Engine(s):

• TensorRT-LLM

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Blackwell

Preferred Operating System(s):

• Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

The model is version 1 of Eagle3 and is trained with nvidia-modelopt v0.44.0

Training and Evaluation Datasets:

** Total size (in number of data points) 112K.

** Dataset partition: Training 100%

Training Dataset:

Link: Nemotron-Post-Training-Dataset-v2, only prompts from the datasets were used for data synthesis, (the original responses from GPT were not used), which is then used to train the Eagle modules. Data Modality: Text, Image, Video

Image Training Data Size: None Text Training Data Size: Less than a Billion Tokens Video Training Data Size: None Data Collection Method by dataset: Hybrid: Automated, Synthetic

Labeling Method by dataset: Hybrid: Automated, Synthetic

Properties: 112K multilingual text samples featuring prompts spanning math, code, STEM, and conversational topics. Each sample includes a synthetic response generated by the target model.

Evaluation Dataset:

Link: MTBench, for more details, see here Data Collection Method by dataset: Hybrid: Human, Synthetic

Labeling Method by dataset: Hybrid: Human, Synthetic Properties: 3,300 multi-turn dialogue sequences, each annotated with expert preference votes.

Link: SPEED-Bench (qualitative subset) Data Collection Method by dataset: Hybrid: Human, Synthetic

Labeling Method by dataset: Hybrid: Human, Synthetic Properties: 880 single- and multi-turn evaluation prompts spanning 11 categories (coding, humanities, math, multilingual, qa, rag, reasoning, roleplay, stem, summarization, writing).

Inference:

Acceleration Engine: TensorRT-LLM

Test Hardware: NVIDIA B200

Eagle Speculative Decoding

Synthesized data was obtained from Moonshot AI's Kimi-K2.6 model, which is then used to finetune the Eagle modules. This model is ready for inference with TensorRT-LLM in Eagle speculative decoding mode. Eagle modules are used to predict candidate tokens beyond the next token. In the generation step, each forward Eagle module generates a distribution of tokens beyond the previous. The longest accepted candidate sequence is selected so that more than 1 token is returned in the generation step. The number of tokens generated in each step is called acceptance rate.

Usage

To serve the checkpoint with TensorRT-LLM, follow the sample commands below with the TensorRT-LLM GitHub repo:

trtllm-serve --host 0.0.0.0 --port 8000 --backend pytorch --max_batch_size 32 --max_num_tokens 8192 --max_seq_len 8192 --tp_size 4 --extra_llm_api_options extra-llm-api-config.yml

with extra-llm-api-config.yml being

speculative_config:
decoding_type: Eagle
max_draft_len: 3
speculative_model_dir:

Evaluation

Acceptance rate on MT-bench with draft length 3:

| Category | MT Bench Acceptance Rate | |-------------|:-----------------------:| | writing | 2.41 | | roleplay | 2.29 | | reasoning | 2.62 | | math | 3.23 | | coding | 2.84 | |…