DeepSeek-V3.2-Exp
DeepSeek's experimental language model — sparse attention, 50%+ cheaper API
DeepSeek's experimental language model with sparse attention and 50%+ cheaper API pricing.
Open foundation models — the base layers everything builds on.
Key features
- Experimental model with DeepSeek Sparse Attention
- Fine-grained sparse attention for long-context training and inference efficiency
- Performance on par with V3.1-Terminus across public benchmarks
- Open-source CUDA kernels and updated demo inference code
- MIT licensed; day-0 support in SGLang and vLLM
Good for
- Long-context research and efficient inference experiments
- Reasoning, coding, and agentic benchmark workloads
Stack & tags
PythonCUDAllmopen-weightssparse-attentionlong-contextmitexperimentalmoe
Quick start
torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --interactiveFrom the project README on GitHub.