DeepSeek: DeepSeek V3.2 Exp

Text input Text output
Author's Description

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.

Key Specifications
Cost
$$$
Context
131K
Released
Sep 29, 2025
Speed
Ability
Reliability
Supported Parameters

This model supports the following parameters:

Logprobs Include Reasoning Temperature Tools Stop Reasoning Max Tokens Presence Penalty Top Logprobs Frequency Penalty Tool Choice Top P
Features

This model supports the following features:

Tools Reasoning
Performance Summary

DeepSeek-V3.2-Exp, an experimental model from DeepSeek, demonstrates moderate speed performance, ranking in the 36th percentile across benchmarks. It offers cost-effective solutions, placing in the 64th percentile for price. Notably, the model exhibits exceptional reliability with a 100% success rate, consistently providing usable responses without technical failures. In terms of benchmark performance, DeepSeek-V3.2-Exp shows strong capabilities across various domains. It achieves high accuracy in Coding (93.0%), General Knowledge (99.5%), and Mathematics (94.0%), indicating robust understanding and problem-solving skills. Its Instruction Following (71.7%) and Reasoning (86.0%) scores are also commendable, suggesting effective processing of complex directives and logical tasks. The model excels in Ethics with a perfect 100% accuracy, making it the most accurate and efficient model at its price point for this category. Hallucinations are well-managed with 96.0% accuracy, and Email Classification is strong at 98.0%. While its primary focus is architectural optimization for long contexts, the model maintains performance comparable to V3.1, with minor trade-offs and gains. Its key strength lies in its high accuracy across critical cognitive tasks and its perfect reliability, making it a valuable research-oriented model for exploring efficient transformer designs.

Model Pricing

Current Pricing

Feature Price (per 1M tokens)
Prompt $0.21
Completion $0.32

Price History

Available Endpoints
Provider Endpoint Name Context Length Pricing (Input) Pricing (Output)
DeepSeek
DeepSeek | deepseek/deepseek-v3.2-exp 131K $0.21 / 1M tokens $0.32 / 1M tokens
Novita
Novita | deepseek/deepseek-v3.2-exp 163K $0.21 / 1M tokens $0.32 / 1M tokens
DeepInfra
DeepInfra | deepseek/deepseek-v3.2-exp 163K $0.21 / 1M tokens $0.32 / 1M tokens
GMICloud
GMICloud | deepseek/deepseek-v3.2-exp 163K $0.21 / 1M tokens $0.32 / 1M tokens
AtlasCloud
AtlasCloud | deepseek/deepseek-v3.2-exp 163K $0.21 / 1M tokens $0.32 / 1M tokens
Novita
Novita | deepseek/deepseek-v3.2-exp 163K $0.27 / 1M tokens $0.41 / 1M tokens
SiliconFlow
SiliconFlow | deepseek/deepseek-v3.2-exp 163K $0.27 / 1M tokens $0.41 / 1M tokens
DeepInfra
DeepInfra | deepseek/deepseek-v3.2-exp 163K $0.21 / 1M tokens $0.32 / 1M tokens
Chutes
Chutes | deepseek/deepseek-v3.2-exp 131K $0.21 / 1M tokens $0.32 / 1M tokens
Benchmark Results
Benchmark Category Reasoning Strategy Free Executions Accuracy Cost Duration
Other Models by deepseek