AngelSlim/Qwen3-32B_eagle3

Dedicated to building a more intuitive, comprehensive, and efficient LLMs compression toolkit.

📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope   |   💬 WeChat

Table of Contents

• Latest Updates
• Key Features
• Supported Models
• How to Use
  • Install AngelSlim
  • Quick Start
  • deployment & Evaluation
• Benchmark
• License
• Citation
• Technical Discussion

📣Latest Updates

• [25/07/04] We now support quantization for Hunyuan/Qwen2.5/Qwen3/DeepSeek-R1-Distill-Qwen and other models, including INT8/FP8/INT4 algorithms. We also opensource Qwen3-8B`s Eagle3 model weight.

Coming soon:

• Support W4A8 quantization for DeepSeek-R1.
• Support quantization for multimodal models like Qwen-VL.
• Release of new algorithm for speculative sampling.

🌟Key Features

• Highly Integrated: This toolkit integrates mainstream compression algorithms into a unified framework, offering developers one-click access with exceptional ease of use.
• Continuous Innovation: Beyond integrating widely-used industry algorithms, we are continuously researching better compression algorithms, which will be gradually open-sourced in the future.
• Performance-Driven: We continuously optimize end-to-end performance in model compression workflows and algorithm deployment, such as enabling quantization of models like Qwen3-235B and DeepSeek-R1 on a single GPU.

💼Supported Models

Quantization

Currently supports the following LLMs, including Hunyuan-Dense, Hunyuan-MoE, Qwen3-Dense, Qwen3-MoE, Qwen2.5, DeepSeek-R1 distilled Qwen models, and QwQ::

Model	FP8-Dynamic	FP8-Static	INT8-Dynamic	INT4-GPTQ	INT4-AWQ
Hunyuan-Dense	✅	✅	✅	✅	✅
Hunyuan-MoE	✅	✅	✅	✅	✅
Qwen3-Dense	✅	✅	✅	✅	✅
Qwen3-MoE	✅	✅	✅	✅	✅
Qwen2.5	✅	✅	✅	✅	✅
DeepSeek-R1-Distill-Qwen	✅	✅	✅	✅	✅
QwQ	✅	✅	✅	✅	✅

Speculative Decoding

The Eagle3 weights for the Qwen3-8B model are now available, with Eagle3 weights for other models in the Qwen3 series to be released soon.

Model	Eagle3
Qwen3-8B	✅
Qwen3-14B	coming soon
Qwen3-32B	coming soon

🛎️How to Use

Install AngelSlim

We recommend using pip to install the latest stable version of AngelSlim:

pip install angelslim

Alternatively, you can clone the repository and install from source in editable mode:

cd AngelSlim && python setup.py install

For more detailed installation instructions, please refer to the Installation Documentation.

Quick Start

After installing AngelSlim, you can quickly start by running the following script to perform static FP8 quantization on the Qwen3-1.7B model:

• One-click Start

  python3 tools/run.py -c configs/qwen3/fp8_static/qwen3-1_7b_fp8_static.yaml
  

  This example will load the HuggingFace model and perform activation value calibration using the dataset specified in the config file, saving the quantized model weights.

• Code-based Start

  To perform dynamic FP8 quantization on Qwen3-1.7B:

  from angelslim.engine import Engine
  
  slim_engine = Engine()
  # Prepare model
  slim_engine.prepare_model(model_name="Qwen", model_path="Qwen/Qwen3-1.7B",)
  # Initialize compressor
  slim_engine.prepare_compressor("PTQ", default_method="fp8_dynamic")
  # Compress model
  slim_engine.run()
  # Save compressed model
  slim_engine.save("./output")
  

For more details, please refer to the Quick Start Documentation.

🖥️ Deployment and Testing

1. API Service Deployment

After specifying the quantized model path MODEL_PATH, you can deploy an OpenAI-compatible API service using the following LLMs inference frameworks:

vLLM

Use the following script to launch a vLLM server, recommended version vllm>=0.8.5.post1. For MOE INT8 quantized models, vllm>=0.9.0 is required.

bash deploy/run_vllm.sh $MODEL_PATH

SGLang

Use the following script to launch a SGLang server, recommended version sglang>=0.4.6.post1.

bash deploy/run_sglang.sh $MODEL_PATH

2. Service Invocation

Invoke requests via OpenAI's API format:

bash deploy/openai.sh $MODEL_PATH

3. Performance Evaluation

Evaluate the performance of quantized model using lm-evaluation-harness, recommended versionlm-eval>=0.4.8:

bash deploy/lm_eval.sh $MODEL_PATH

For more detaileds, please refer to the Deployment Documentation.

📈 Benchmark

Quantization

The performance test results for selected models are shown below. For the complete benchmark, refer to the Benchmark documentation

Hunyuan Series Models

Benchmark results for the Hunyuan-A13B-Instruct model with FP8 and INT4-GPTQ quantization algorithms on datasets including AIME 2024, GSM8K, BBH, and DROP:

Bench	Hunyuan-A13B-Instruct	Hunyuan-A13B-Instruct-FP8	Hunyuan-A13B-Instruct-Int4-GPTQ
AIME 2024	87.3	86.7	86.7
GSM8K	94.39	94.01	94.24
BBH	89.1	88.34	87.91
DROP	91.1	91.1	91.05

Qwen3 Series Models

Benchmark results for Qwen3 series models with FP8-Static, FP8-Dynamic, INT4-GPTQ, and INT4-AWQ quantization algorithms on datasets including CEVAL, MMLU, GSM8K, and HUMANEVAL:

Model	Quantization	CEVAL	MMLU	GSM8K	HUMANEVAL
Qwen3-0.6B	BF16	45.84	47.21	42.99	19.51
	FP8-Static	45.99	46.87	38.06	18.90
	FP8-Dynamic	45.99	46.93	38.29	20.73
	INT8-Dynamic	45.17	46.95	41.17	21.34
Qwen3-8B	BF16	79.27	74.78	87.79	63.41
	FP8-Static	78.23	74.79	86.96	62.20
	FP8-Dynamic	78.45	74.75	87.64	62.80
	INT8-Dynamic	78.01	74.84	86.96	67.07
	INT4-GPTQ	77.19	73.26	86.43	62.20
	INT4-AWQ	76.15	73.59	86.96	63.41
Qwen3-14B	BF16	83.06	78.90	88.40	55.49
	FP8-Static	82.62	78.57	89.46	57.32
	FP8-Dynamic	82.24	78.92	88.32	52.44
	INT8-Dynamic	81.87	78.13	86.28	56.10
	INT4-GPTQ	81.05	78.02	87.34	57.93
	INT4-AWQ	82.02	77.68	84.23	61.59
Qwen3-32B	BF16	86.55	82.00	74.53	37.80
	FP8-Static	86.92	81.78	70.20	39.63
	FP8-Dynamic	86.55	81.89	70.43	38.41
	INT4-GPTQ	86.18	81.01	-	43.29
	INT4-AWQ	86.18	81.54	-	36.59
Qwen3-30B-A3B	BF16	83.66	79.36	89.99	31.71
	FP8-Static	83.95	79.47	89.01	31.10
	FP8-Dynamic	84.10	79.40	89.16	32.93
	INT8-Dynamic	83.36	79.48	89.16	34.15
Qwen3-235B-A22B	BF16	89.60	86.28	85.29	27.44
	FP8-Static	89.67	86.19	86.96	27.44
	FP8-Dynamic	89.67	86.18	85.22	28.05
	INT8-Dynamic	88.93	86.20	86.20	23.78
QwQ-32B	BF16	85.74	82.03	73.31	42.68
	FP8-Static	85.44	81.91	75.36	42.68
	FP8-Dynamic	85.07	81.93	75.66	42.07
	INT4-GPTQ	84.03	81.26	68.23	45.73
	INT4-AWQ	83.58	81.01	68.69	43.29

Other Models

Benchmark results for other models with FP8-Static, FP8-Dynamic, INT4-GPTQ, and INT4-AWQ quantization algorithms on datasets including CEVAL, MMLU and GSM8K:

Model	Quantization	CEVAL	MMLU	GSM8K
Qwen2.5-1.5B-Instruct	BF16	67.01	60.05	54.28
	FP8-Static	66.27	60.23	-
	FP8-Dynamic	66.79	60.08	51.71
Qwen2.5-7B-Instruct	BF16	81.20	74.55	79.98
	FP8-Static	81.13	74.03	79.30
	FP8-Dynamic	80.31	74.07	79.00
	INT4-GPTQ	79.05	73.05	74.75
	INT4-AWQ	79.35	73.22	79.38
Qwen2.5-32B-Instruct	BF16	87.30	83.21	81.73
	FP8-Static	87.59	83.08	81.58
	FP8-Dynamic	87.30	83.04	81.58
	INT4-GPTQ	86.70	82.45	82.03
	INT4-AWQ	87.00	82.64	-
DeepSeek-R1-Distill-Qwen-7B	BF16	53.49	53.80	75.74
	FP8-Static	53.57	54.17	76.19
	FP8-Dynamic	52.97	54.13	74.15
	INT4-GPTQ	51.86	52.44	75.89
	INT4-AWQ	53.49	53.70	-
DeepSeek-R1-Distill-Qwen-14B	BF16	77.71	74.28	85.67
	FP8-Static	77.56	74.66	86.73
	FP8-Dynamic	76.82	74.63	87.11
	INT4-GPTQ	74.29	72.37	84.61
	INT4-AWQ	74.81	73.00	86.05
DeepSeek-R1-Distill-Qwen-32B	BF16	84.18	80.89	87.41
	FP8-Static	83.43	80.90	87.57
	FP8-Dynamic	83.73	81.10	86.43
	INT4-GPTQ	84.10	79.80	86.73
	INT4-AWQ	82.84	80.15	87.19

Speculative Decoding

Benchmark results for Qwen3 series models with Eagle3 speculative decoding algorithm on datasets including MT-bench, HunmanEval, GSM8K, and Alpaca:

Qwen3-8B

Temperature	Method	Datasets
		MT-bench		HumanEval		GSM8K		Alpaca
		Speedup	Accept length	Speedup	Accept length	Speedup	Accept length	Speedup	Accept length
T=0	Eagle3	2.63x	3.65	2.76x	3.85	2.82x	3.90	2.62x	3.48
T=1	Eagle3	1.98x	2.75	2.25x	3.11	2.31x	3.15	2.10x	2.76

📝 Model License

The code for this project is open-sourced under the License for AngelSlim.

🔗 Citation

@software{AngelSlim2025,
    title={{AngelSlim}},
    author={Tencent AngelSlim Project Contributors},
    year={2025},
    month={6},
    url={https://github.com/Tencent/AngelSlim},
}

💬 Technical Discussion

• AngelSlim is continuously iterating and new features will be released soon. If you have any questions or suggestions, please open an issue on GitHub or join our WeChat technical discussion group.