For more details, please refer to vLLM documentation [link]
【vLLM Startup Command - Single Node with 6 GPUs】
Note, as of 2025-09-30:
Only Hopper and Blackwell data center GPUs are supported for now.
The kernels are mainly optimized for TP=1, so it is recommended to run this model under DP/EP mode
DP mode may result in increased serving latency. To mitigate this, we recommend enabling MTP to maintain optimal speed. To disable MTP, simply remove the --speculative-config flag.
Some users have observed improved performance on H20 machines by setting export VLLM_USE_DEEP_GEMM=0
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
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:
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.
