DeepSeek V3.2 Exp

DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism designed to improve training and inference efficiency in long-context scenarios while maintaining output quality. Users can control the reasoning behaviour with the `reasoning` `enabled` boolean. [Learn more in our docs](https://openrouter.ai/docs/use-cases/reasoning-tokens#enable-reasoning-with-default-config) The model was trained under conditions aligned with V3.1-Terminus to enable direct comparison. Benchmarking shows performance roughly on par with V3.1 across reasoning, coding, and agentic tool-use tasks, with minor tradeoffs and gains depending on the domain. This release focuses on validating architectural optimizations for extended context lengths rather than advancing raw task accuracy, making it primarily a research-oriented model for exploring efficient transformer designs.

Licence

mit

Architecture

DeepseekV32ForCausalLM

Parameters

685B

Context

163.8K tokens

Artifact format

safetensors

Quantization

fp8 8-bit

Download size

642 GB

Hardware to run it (inference, estimated)

766 GB VRAM minimum · 958 GB recommended · 958 GB system RAM · 643 GB storage

Estimated from parameter count and stored precision; verify against the selected runtime and context length.

Benchmarks

No benchmark observation recorded for this model.