Hugging Face's logo

deepseek-ai
/
DeepSeek-V3.2-Exp

Text Generation
Transformers
Safetensors
deepseek_v32
conversational
fp8
Model card Files
xet
Community
33
DeepSeek-V3.2-Exp / README.md
mackenzietechdocs's picture
mackenzietechdocs
Added Evaluation Benchmarks to Metadata
6660ced verified
preview code
|
raw
history blame
8.26 kB
metadata
license: mit
library_name: transformers
base_model:
  - deepseek-ai/DeepSeek-V3.2-Exp-Base
base_model_relation: finetune
model-index:
  - name: DeepSeek-V3.2-Exp
    results:
      - task:
          type: text-generation
        dataset:
          name: Benchmarks
          type: benchmark
        metrics:
          - name: MMLU-Pro
            type: mmlu-pro
            value: 85
          - name: GPQA-Diamond
            type: gpqa-diamond
            value: 79.9
          - name: Humanity's Last Exam
            type: humanity's_last_exam
            value: 19.8
          - name: LiveCodeBench
            type: livecodebench
            value: 74.1
          - name: AIME 2025
            type: aime_2025
            value: 89.3
          - name: HMMT 2025
            type: hmmt_2025
            value: 83.6
          - name: Codeforces
            type: codeforces
            value: 2121
          - name: Aider-Polyglot
            type: aider-polyglot
            value: 74.5
          - name: BrowseComp
            type: browsecomp
            value: 40.1
          - name: BrowseComp-zh
            type: browsecomp-zh
            value: 47.9
          - name: SimpleQA
            type: simpleqa
            value: 97.1
          - name: SWE Verified
            type: swe_verified
            value: 67.8
          - name: SWE-bench Multilingual
            type: swe-bench_multilingual
            value: 57.9
          - name: Terminal-bench
            type: terminal-bench
            value: 37.7
        source:
          name: Model README
          url: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp

DeepSeek-V3

Introduction

We are excited to announce the official release of DeepSeek-V3.2-Exp, an experimental version of our model. As an intermediate step toward our next-generation architecture, V3.2-Exp builds upon V3.1-Terminus by introducing DeepSeek Sparse Attention—a sparse attention mechanism designed to explore and validate optimizations for training and inference efficiency in long-context scenarios.

This experimental release represents our ongoing research into more efficient transformer architectures, particularly focusing on improving computational efficiency when processing extended text sequences.

  • DeepSeek Sparse Attention (DSA) achieves fine-grained sparse attention for the first time, delivering substantial improvements in long-context training and inference efficiency while maintaining virtually identical model output quality.

  • To rigorously evaluate the impact of introducing sparse attention, we deliberately aligned the training configurations of DeepSeek-V3.2-Exp with V3.1-Terminus. Across public benchmarks in various domains, DeepSeek-V3.2-Exp demonstrates performance on par with V3.1-Terminus.

Benchmark DeepSeek-V3.1-Terminus DeepSeek-V3.2-Exp
Reasoning Mode w/o Tool Use
MMLU-Pro 85.0 85.0
GPQA-Diamond 80.7 79.9
Humanity's Last Exam 21.7 19.8
LiveCodeBench 74.9 74.1
AIME 2025 88.4 89.3
HMMT 2025 86.1 83.6
Codeforces 2046 2121
Aider-Polyglot 76.1 74.5
Agentic Tool Use
BrowseComp 38.5 40.1
BrowseComp-zh 45.0 47.9
SimpleQA 96.8 97.1
SWE Verified 68.4 67.8
SWE-bench Multilingual 57.8 57.9
Terminal-bench 36.7 37.7

Update

  • 2025.11.17: We have identified that previous versions of the inference demo code contained an implementation discrepancy in Rotary Position Embedding (RoPE) within the indexer module, potentially leading to degraded model performance. Specifically, the input tensor to RoPE in the indexer module requires a non-interleaved layout, whereas RoPE in the MLA module expects an interleaved layout. This issue has now been resolved. Please refer to the updated version of the inference demo code and take note of this implementation detail.

How to Run Locally

HuggingFace

We provide an updated inference demo code in the inference folder to help the community quickly get started with our model and understand its architectural details.

First convert huggingface model weights to the the format required by our inference demo. Set MP to match your available GPU count:

cd inference
export EXPERTS=256
python convert.py --hf-ckpt-path ${HF_CKPT_PATH} --save-path ${SAVE_PATH} --n-experts ${EXPERTS} --model-parallel ${MP}

Launch the interactive chat interface and start exploring DeepSeek's capabilities:

export CONFIG=config_671B_v3.2.json
torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --interactive

SGLang

Installation with Docker 

# H200
docker pull lmsysorg/sglang:dsv32

# MI350
docker pull lmsysorg/sglang:dsv32-rocm

# NPUs
docker pull lmsysorg/sglang:dsv32-a2
docker pull lmsysorg/sglang:dsv32-a3

Launch Command 

python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention

vLLM

vLLM provides day-0 support of DeepSeek-V3.2-Exp. See the recipes for up-to-date details.

Open-Source Kernels

For TileLang kernels with better readability and research-purpose design, please refer to TileLang.

For high-performance CUDA kernels, indexer logit kernels (including paged versions) are available in DeepGEMM. Sparse attention kernels are released in FlashMLA.

License

This repository and the model weights are licensed under the MIT License.

Citation 

@misc{deepseekai2024deepseekv32,
      title={DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention}, 
      author={DeepSeek-AI},
      year={2025},
}

Contact

If you have any questions, please raise an issue or contact us at service@deepseek.com.