DeepSeek

DeepSeek V3.2

by DeepSeektext->text14 endpoints
Neura Intelligence Index
67.2/ 100
Rank
#91
Confidence
medium
Benchmarks
6(15% coverage)

Overview

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...

Capabilities

Text generation
Tool use / Function calling
Structured output
Extended reasoning
Prompt caching
Long context
Top-P sampling
Stop sequences
Deterministic seed
Log probabilities

Modalities

Input
Text
Output
Text

Technical Specifications

Context Window
163.8K tokens
Max Output
147.5K tokens
Tokenizer
DeepSeek
Uptime (24h)
10000.0%
Parameters
685000000000B
License
mitOpen Weights
Country
🇨🇳CN
Release Date
December 1, 20259mo

Supported Parameters

frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p

Benchmark Performance

coding

BenchmarkScoreSource
SWE-bench Verified
73.1
verified

math

BenchmarkScoreSource
AIME 2025
93.1
verified

reasoning

BenchmarkScoreSource
GPQA Diamond
82.4
verified
Humanity's Last Exam
40.8
verified

agent

BenchmarkScoreSource
BrowseComp
51.4
verified
Toolathlon
35.2
verified

Performance Over Time

Pricing

Per 1M tokens
Input$0.26
Output$0.38
Cache read$0.13
Added
December 1, 2025
Last synced 8/27/2026