🤗 HuggingFace |  🤖 ModelScope |  🪡 AngelSlim

🖥️ Official Website  |   🕖 HunyuanAPI  |   🕹️ Demo    

GITHUB | cnb.cool | LICENSE | WeChat | Discord

Model Introduction

Hunyuan is Tencent’s open-source efficient large language model series, designed for versatile deployment across diverse computational environments. From edge devices to high-concurrency production systems, these models deliver optimal performance with advanced quantization support and ultra-long context capabilities.

We have released a series of Hunyuan dense models, comprising both pre-trained and instruction-tuned variants, with parameter scales of 0.5B, 1.8B, 4B, and 7B. These models adopt training strategies similar to the Hunyuan-A13B, thereby inheriting its robust performance characteristics. This comprehensive model family enables flexible deployment optimization - from resource-constrained edge computing with smaller variants to high-throughput production environments with larger models, all while maintaining strong capabilities across diverse scenarios.

Key Features and Advantages

• Hybrid Reasoning Support: Supports both fast and slow thinking modes, allowing users to flexibly choose according to their needs.
• Ultra-Long Context Understanding: Natively supports a 256K context window, maintaining stable performance on long-text tasks.
• Enhanced Agent Capabilities: Optimized for agent tasks, achieving leading results on benchmarks such as BFCL-v3, τ-Bench and C3-Bench.
• Efficient Inference: Utilizes Grouped Query Attention (GQA) and supports multiple quantization formats, enabling highly efficient inference.

Related News

• 2025.7.30 We have open-sourced Hunyuan-0.5B-Pretrain , Hunyuan-0.5B-Instruct , Hunyuan-1.8B-Pretrain , Hunyuan-1.8B-Instruct , Hunyuan-4B-Pretrain , Hunyuan-4B-Instruct , Hunyuan-7B-Pretrain ,Hunyuan-7B-Instruct on Hugging Face.
  

Benchmark

Note: The following benchmarks are evaluated by TRT-LLM-backend on several base models.

Use with transformers

First, please install transformers. We will merge it into the main branch later.

pip install git+https://github.com/huggingface/transformers@4970b23cedaf745f963779b4eae68da281e8c6ca

Our model defaults to using slow-thinking reasoning, and there are two ways to disable CoT reasoning. 1. Pass “enable_thinking=False” when calling apply_chat_template. 2. Adding “/no_think” before the prompt will force the model not to use perform CoT reasoning. Similarly, adding “/think” before the prompt will force the model to perform CoT reasoning.

The following code snippet shows how to use the transformers library to load and apply the model. It also demonstrates how to enable and disable the reasoning mode , and how to parse the reasoning process along with the final output.

we use tencent/Hunyuan-7B-Instruct for example

from transformers import AutoModelForCausalLM, AutoTokenizer
import os
import re

model_name_or_path = "tencent/Hunyuan-7B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, device_map="auto")  # You may want to use bfloat16 and/or move to GPU here
messages = [
    {"role": "user", "content": "Write a short summary of the benefits of regular exercise"},
]
tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True,return_tensors="pt",
                                                enable_thinking=True # Toggle thinking mode (default: True)
                                                )

outputs = model.generate(tokenized_chat.to(model.device), max_new_tokens=2048)

output_text = tokenizer.decode(outputs[0])
print("output_text=",output_text)
think_pattern = r'<think>(.*?)</think>'
think_matches = re.findall(think_pattern, output_text, re.DOTALL)

answer_pattern = r'<answer>(.*?)</answer>'
answer_matches = re.findall(answer_pattern, output_text, re.DOTALL)

think_content = [match.strip() for match in think_matches][0]
answer_content = [match.strip() for match in answer_matches][0]
print(f"thinking_content:{think_content}\n\n")
print(f"answer_content:{answer_content}\n\n")

We recommend using the following set of parameters for inference. Note that our model does not have the default system_prompt.

{
  "do_sample": true,
  "top_k": 20,
  "top_p": 0.8,
  "repetition_penalty": 1.05,
  "temperature": 0.7
}

Training Data Format

If you need to fine-tune our Instruct model, we recommend processing the data into the following format, corresponding to both slow-thinking and fast-thinking scenarios.

# think_pattern
think = ""
answer = ""
think_pattern = f"<think>\n{think}\n</think>\n<answer>\n{answer}\n</answer>"

# fast think pattern
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "/no_think海水为什么是咸的" },
    {"role": "assistant", "content": "<think>\n\n</think>\n<answer>\n海水是咸的主要是因为其中含有许多溶解在水中的盐类和矿物质。这些盐类和矿物质来自于地球表面的岩石和土壤中的化学物质，随着时间的推移，它们被带到了海洋中。当海水蒸发时，水分蒸发掉了，但盐类和矿物质仍然留在水中，导致海水变得更加咸味。因此，海水的咸度是由其中的盐类和矿物质的含量决定的。\n</answer>"}
]

# slow think pattern
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "1+1=" },
    {"role": "assistant", "content": "<think>\n嗯，用户问的是1加1等于多少。首先，我需要确认这是一个基本的算术问题。1加1在十进制的数学体系中，通常的结果是2。不过，可能需要考虑是否有其他情况，比如二进制或者其他数制，但用户没有特别说明，所以默认应该是十进制。另外，有时候可能会有脑筋急转弯的情况，比如在某些语境下1+1可能等于1（比如1滴水加1滴水还是1滴水），但通常数学问题中都是2。所以最准确的回答应该是2。</think>\n<answer>\n在十进制的基本算术运算中，1加1的结果是2。这是数学中最基础的加法运算之一，遵循自然数的加法规则。因此，1 + 1 = 2。\n</answer>"}
]

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("your_tokenizer_path", trust_remote_code=True)
train_ids = tokenizer.apply_chat_template(messages)

Train with LLaMA-Factory

In the following chapter, we will introduce how to use LLaMA-Factory to fine-tune the Hunyuan model.

Prerequisites

Verify installation of the following dependencies:
- LLaMA-Factory: Follow official installation guide - DeepSpeed (optional): Follow official installation guide - Transformer Library: Use the companion branch (Hunyuan-submitted code is pending review) pip install git+https://github.com/huggingface/transformers@4970b23cedaf745f963779b4eae68da281e8c6ca

Data preparation

We need to prepare a custom dataset: 1. Organize your data in json format and place it in the data directory in LLaMA-Factory. The current implementation uses the sharegpt dataset format, which requires the following structure:

[
  {
    "messages": [
      {
        "role": "system",
        "content": "System prompt (optional)"
      },
      {
        "role": "user",
        "content": "Human instruction"
      },
      {
        "role": "assistant",
        "content": "Model response"
      }
    ]
  }
]

Refer to the Data Format section mentioned earlier for details.

1. Define your dataset in the data/dataset_info.json file using the following format:

"dataset_name": {
  "file_name": "dataset.json",
  "formatting": "sharegpt",
  "columns": {
    "messages": "messages"
  },
  "tags": {
    "role_tag": "role",
    "content_tag": "content",
    "user_tag": "user",
    "assistant_tag": "assistant",
    "system_tag": "system"
  }
}

Training execution

1. Copy all files from the train/llama_factory_support/example_configs directory to the example/hunyuan directory in LLaMA-Factory.
2. Modify the model path and dataset name in the configuration file hunyuan_full.yaml. Adjust other configurations as needed:

### model
model_name_or_path: [!!!add the model path here!!!]

### dataset
dataset: [!!!add the dataset name here!!!]

1. Execute training commands: ​​Single-node training​​ Note: Set the environment variable DISABLE_VERSION_CHECK to 1 to avoid version conflicts. export DISABLE_VERSION_CHECK=1 llamafactory-cli train examples/hunyuan/hunyuan_full.yaml Multi-node training​​ Execute the following command on each node. Configure NNODES, NODE_RANK, MASTER_ADDR, and MASTER_PORT according to your environment: export DISABLE_VERSION_CHECK=1 FORCE_TORCHRUN=1 NNODES=${NNODES} NODE_RANK=${NODE_RANK} MASTER_ADDR=${MASTER_ADDR} MASTER_PORT=${MASTER_PORT} \ llamafactory-cli train examples/hunyuan/hunyuan_full.yaml

Quantization Compression

We used our own AngleSlim compression tool to produce FP8 and INT4 quantization models. AngleSlim is a toolset dedicated to creating a more user-friendly, comprehensive and efficient model compression solution.

FP8 Quantization

We use FP8-static quantization, FP8 quantization adopts 8-bit floating point format, through a small amount of calibration data (without training) to pre-determine the quantization scale, the model weights and activation values will be converted to FP8 format, to improve the inference efficiency and reduce the deployment threshold. We you can use AngleSlim quantization, you can also directly download our quantization completed open source model to use LINK.

Int4 Quantization

We use the GPTQ and AWQ algorithm to achieve W4A16 quantization.

GPTQ processes the model weights layer by layer, uses a small amount of calibration data to minimize the reconfiguration error of the quantized weights, and adjusts the weights layer by layer by the optimization process of approximating the Hessian inverse matrix. The process eliminates the need to retrain the model and requires only a small amount of calibration data to quantize the weights, improving inference efficiency and lowering the deployment threshold. AWQ using a small amount of calibration data (without the need for training), the amplitude of the activation values is statistically calculated. For each weight channel, a scaling coefficient s is computed to expand the numerical range of important weights, allowing more information to be retained during quantization.

You can use AngleSlim quantization, you can also directly download our quantization completed open source model to use LINK.

Quantization Benchmark

This subsection describes the Benchmark metrics for the Hunyuan quantitative model.

Deployment

For deployment, you can use frameworks such as TensorRT-LLM, vLLM, or SGLang to serve the model and create an OpenAI-compatible API endpoint.

image: https://hub.docker.com/r/hunyuaninfer/hunyuan-7B/tags

Contact Us

If you would like to leave a message for our R&D and product teams, Welcome to contact our open-source team . You can also contact us via email (hunyuan_opensource@tencent.com).