DeepSeek-v3.2-Exp (huggingface.co)

DeepSeek-AI released DeepSeek-V3.2-Exp, an experimental intermediate model that integrates DeepSeek Sparse Attention (DSA) to explore efficiency optimizations for long-context transformer workloads. The team kept training configurations identical to the prior V3.1-Terminus to isolate the impact of DSA; across public benchmarks the experimental model delivers virtually identical output quality while reporting “substantial” gains in long-context training and inference efficiency. Notable benchmark parity includes identical MMLU-Pro (85.0), a small win on AIME 2025 (89.3 vs 88.4), and a higher Codeforces score (2121 vs 2046), alongside a few modest regressions in specialized math contests—illustrating that sparse attention can be swapped in with minimal quality trade-offs. Technically, DSA implements fine-grained sparse attention to reduce compute and memory when processing extended sequences, a practical testbed for next-gen transformer architectures. The release includes conversion and inference demo code (instructions for model-parallel setup, expert counts, and interactive launch), day‑0 vLLM support, TileLang kernels for readable research implementations, and Docker images for H200/MI350/NPUs. The repo and weights are MIT-licensed. For researchers and deployers, this offers a ready way to evaluate sparse-attention impacts on long-context efficiency without retraining from scratch, and a pathway toward more compute- and memory-efficient large models.