Hunyuan 7B-Pretrain Open-source Large Language Model - Bilingual Communication in Chinese and English, Powerful Performance and Free Deployment

Hunyuan 7B Pretrain

Developed by tencent

Hunyuan 7B is an open-source bilingual large language model (Chinese and English) by Tencent, featuring optimized data ratios and training methods for robust performance, making it one of the strongest Chinese 7B Dense models available.

Large Language Model

Transformers

EnglishOpen Source License:Other #256K long text processing #Best performance in Chinese 7B #GQA attention mechanism

Downloads 56

Release Time : 1/24/2025

Model Overview

Hunyuan 7B is an open-source large language model with 7B parameters, supporting both Chinese and English, and excels in text generation and comprehension.

Model Features

Extended long-text capability

Supports context lengths up to 256K, ideal for processing long documents and complex tasks.

Efficient inference framework

Offers two inference backends: vLLM and TensorRT-LLM, optimizing inference performance.

Outstanding Chinese performance

Excels in multiple Chinese benchmarks, ranking among the strongest Chinese 7B models.

Model Capabilities

Text generation

Text comprehension

Question answering

Code generation

Mathematical reasoning

Use Cases

Natural language processing

Intelligent Q&A

Used to build knowledge-based question-answering systems to address various user queries.

Performs excellently on benchmarks like MMLU

Content creation

Assists in generating long-form content such as articles and reports.

Supports 256K long-text processing

Education

Math problem-solving

Helps students solve math problems and provides step-by-step solutions.

Performs excellently on GSM8K and MATH benchmarks

🚀 Hunyuan-7B Models

The 7B models released by Hunyuan offer strong performance and achieve a good balance between computing and performance.

🚀 Quick Start

You can refer to the content in Tencent-Hunyuan-Large to get started quickly. The training and inference code can use the version provided in this github repository.

Inference Framework

This open-source release offers two inference backend options tailored for the Hunyuan-7B model: the popular vLLM-backend and the TensorRT-LLM Backend. In this release, we are initially open-sourcing the vLLM solution, with plans to release the TRT-LLM solution in the near future.

✨ Features

The 7B models released by Hunyuan this time: Hunyuan-7B-Pretrain and Hunyuan-7B-Instruct , use better data allocation and training, have strong performance, and have achieved a good balance between computing and performance. It stands out from many large-scale language models and is currently one of the strongest Chinese 7B Dense models.

Introduction to Technical Advantages

Model

Extended long text capability to 256K and utilizes Grouped Query Attention (GQA)

Inference Framework

This open-source release offers two inference backend options tailored for the Hunyuan-7B model: the popular vLLM-backend and the TensorRT-LLM Backend. In this release, we are initially open-sourcing the vLLM solution, with plans to release the TRT-LLM solution in the near future.

Training Framework

The Hunyuan-7B open-source model is fully compatible with the Hugging Face format, enabling researchers and developers to perform model fine-tuning using the hf-deepspeed framework. Learn more : Tencent-Hunyuan-Large 。

📚 Documentation

Benchmark

Note: The following benchmarks are evaluated by TRT-LLM-backend

Hunyuan-7B-Pretrain

Property	Qwen2.5-7B	Llama3-8B	OLMO2-7B	HunYuan-7B-V2
MMLU	74.26	66.95	63.7	75.37
MMLU-Pro	46.17	34.04	31	47.54
MMLU-CF	61.01	55.21	52.94	59.62
MMLU-Redux	73.47	66.44	63.74	74.54
BBH	70.4	62.16	38.01	70.77
HellaSwag	75.82	78.24	61.97	80.77
WinoGrande	69.69	73.64	74.43	71.51
PIQA	79.33	80.52	80.63	81.45
SIQA	77.48	61.05	65.2	79.73
NaturalQuestions	31.77	35.43	36.9	33.52
DROP	68.2	60.13	60.8	68.63
ARC-C	91.64	77.59	74.92	91.97
TriviaQA	69.31	78.61	78	74.31
Chinese-SimpleQA	30.37	19.4	7.35	30.51
SimpleQA	4.98	7.68	4.51	3.73
CMMLU	81.39	50.25	38.79	82.19
C-Eval	81.11	50.4	38.53	82.12
C3	71.77	61.5	54	79.07
GSM8K	82.71	57.54	67.5	93.33
MATH	49.6	18.45	19	62.15
CMATH	84.33	52.83	44	88.5
HumanEval	57.93	35.98	15.24	59.15

Hunyuan-7B-Instruct

Property	Qwen2.5-7B-Instruct	Llama-3-8B-Instruct	OLMo-2-1124-7B-DPO	Hunyuan-7B-Instruct
ARC-C	89.83	82.4	-	88.81
BBH	66.24	-	46.6	76.47
CEval	76.82	-	-	81.8
CMMLU	78.55	-	-	82.29
DROP_F1	80.63	-	60.5	82.96
GPQA	36.87	34.6	-	47.98
Gsm8k	80.14	80.6	85.1	90.14
HellaSwag	83.34	-	-	86.57
HumanEval	84.8	60.4	-	84.0
MATH	72.86	-	32.5	70.64
MMLU	72.36	68.5	61.3	79.18

Inference Performance

This section presents the efficiency test results of deploying various models using vLLM, including inference speed (tokens/s) under different batch sizes.

Inference Framework	Model	Number of GPUs (GPU productA)	input_length	batch=1	batch=4
vLLM	hunyuan-7B	1	2048	78.9	279.5

📄 License

The model uses the Tencent License.

📞 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).

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご

Hunyuan 7B Pretrain

Model Overview

Model Features

Model Capabilities

Use Cases

🚀 Hunyuan-7B Models

🚀 Quick Start

Inference Framework

✨ Features

Introduction to Technical Advantages

Model

Inference Framework

Training Framework

📚 Documentation

Related News

Benchmark

Inference Performance

📄 License

📞 Contact Us