DeepSeek V3.2-Exp is completely free with 5 features included. No paid tiers offered, making it perfect for budget-conscious users.
DeepSeek Sparse Attention (DSA) is a fine-grained sparse attention mechanism introduced in V3.2-Exp that replaces the dense attention used in V3.1-Terminus. It delivers substantial improvements in long-context training and inference efficiency while maintaining virtually identical model output quality. For teams processing long documents, codebases, or extended agent traces, this translates directly into lower GPU memory pressure and faster throughput. According to DeepSeek, this is the first time fine-grained sparse attention has been achieved at this scale.
The model weights and repository are released under the MIT License, meaning the model itself is free to download, modify, and deploy commercially. The actual cost is the GPU infrastructure required to serve it — the 671B-parameter MoE typically runs with tensor parallelism of 8 across high-memory GPUs like the H200. Compared to per-token API pricing from closed-weight competitors, self-hosting V3.2-Exp can dramatically reduce inference costs at scale, but small-volume users may find third-party hosted inference providers more economical.
DeepSeek officially provides Docker images targeting NVIDIA H200 GPUs, AMD MI350 accelerators, and Ascend NPUs (A2 and A3 variants). The recommended SGLang launch configuration uses tensor parallelism of 8 with data parallelism of 8 and DP attention enabled. Practically, this means an 8-GPU node with high-bandwidth memory is the minimum reasonable deployment target. Quantized variants distributed by the community via llama.cpp, Ollama, and LM Studio can lower the bar, though with quality and context-length tradeoffs.
DeepSeek deliberately aligned the training configurations of the two models to isolate the effect of sparse attention. Results are essentially a wash with small movements in either direction: MMLU-Pro is identical at 85.0, AIME 2025 improves to 89.3 (from 88.4), Codeforces rating rises to 2121 (from 2046), and SimpleQA edges up to 97.1. Slight regressions appear on GPQA-Diamond (79.9 vs 80.7) and Humanity's Last Exam (19.8 vs 21.7). The point of the release is the efficiency win from DSA, not benchmark improvements.
DeepSeek explicitly labels this as an experimental release intended to validate optimizations for the next-generation architecture, not as a stable production model. A notable RoPE implementation bug in the indexer module was identified and patched on 2025-11-17, which is the type of rough edge typical of research releases. Teams that need production stability should weigh whether to wait for the non-experimental successor or to pin a specific commit and validate thoroughly. For research, evaluation, and internal tooling the MIT license and benchmark parity make it an attractive choice.
It's completely free — no credit card required.
Start Using DeepSeek V3.2-Exp — It's Free →Still not sure? Read our full verdict →
Last verified March 2026