Model Introduction
The 7B models released by Hunyuan this time: Hunyuan-7B-Pretrain-0124 and Hunyuan-7B-Instruct-0124 , 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 。
Related News
- 2025.1.24 We have open-sourced Hunyuan-7B-Pretrain-0124 , Hunyuan-7B-Instruct-0124 on Hugging Face.
Benchmark
Note: The following benchmarks are evaluated by TRT-LLM-backend
Hunyuan-7B-Pretrain
| 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
| Model | 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 |
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.
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 |
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).