DeepSeek-V3.2 (Thinking)

Name: DeepSeek-V3.2 (Thinking)
Author: DeepSeek

DeepSeek

DeepSeek-V3.2 in thinking mode. A powerful 685B parameter reasoning model using DeepSeek Sparse Attention (DSA). This mode enables extended chain-of-thought reasoning for solving complex tasks. Supports JSON output and tool calls.

Key Specifications

Parameters

685.0B

Context

131.1K

Release Date

November 30, 2025

Average Score

68.5%

API Documentation Research Paper Model Weights

Timeline

Key dates in the model's history

Announcement

November 30, 2025

Last Update

January 22, 2026

Today

September 10, 2026

Technical Specifications

Parameters

685.0B

Training Tokens

Knowledge Cutoff

Family

Capabilities

MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)

$0.28

Output (per 1M tokens)

$0.42

Max Input Tokens

131.1K

Max Output Tokens

65.5K

Supported Features

Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Programming

Programming skills tests

SWE-Bench Verified

Resolved • Self-reported

73.1%

Reasoning

Logical reasoning and analysis

GPQA

Diamond subset, Pass@1 • Self-reported

82.0%

Other Tests

Specialized benchmarks

AIME 2025

Pass@1 • Self-reported

93.0%

BrowseComp

Pass@1 • Self-reported

51.4%

BrowseComp-zh

Pass@1 • Self-reported

65.0%

CodeForces

Rating • Self-reported

79.5%

HMMT 2025

Pass@1 • Self-reported

90.0%

Humanity's Last Exam

Pass@1 • Self-reported

25.1%

LiveCodeBench

Pass@1-COT • Self-reported

83.0%

MMLU-Pro

EM • Self-reported

85.0%

SWE-bench Multilingual

Resolved • Self-reported

70.2%

T2-Bench

Pass@1 • Self-reported

80.0%

Terminal-Bench 2.0

Acc • Self-reported

46.4%

Toolathlon

Pass@1 • Self-reported

35.2%

License & Metadata

License

mit

Announcement Date

November 30, 2025

Last Updated

January 22, 2026

Articles about DeepSeek-V3.2 (Thinking)

The DeepSeek V4 'Leak' Was Fake. But the Real Model May Be Bigger Than Anyone Expected.

A viral Reddit post about a massive new DeepSeek model turned out to be fabricated. The actual V4 — ~1 trillion parameters, 1M context — is still coming.

March 27, 2026

3 min

Similar Models

All Models

DeepSeek-V3.1

DeepSeek

671.0B

Best score:0.8 (GPQA)

Released:Jan 2025

Price:$0.27/1M tokens

DeepSeek-V4-Flash-0423

DeepSeek

284.0B

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$0.14/1M tokens

DeepSeek-V3

DeepSeek

671.0B

Best score:0.9 (MMLU)

Released:Dec 2024

Price:$0.27/1M tokens

DeepSeek-R1

DeepSeek

671.0B

Best score:0.9 (MMLU)

Released:Jan 2025

Price:$3.00/1M tokens

DeepSeek-V4-Flash-Max

DeepSeek

284.0B

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$0.14/1M tokens

DeepSeek-V4-Pro-Max

DeepSeek

1.6T

Best score:0.9 (GPQA)

Released:Apr 2026

Price:$1.74/1M tokens

DeepSeek-V2.5

DeepSeek

236.0B

Best score:0.9 (HumanEval)

Released:May 2024

Price:$2.00/1M tokens

DeepSeek-V3.2-Exp

DeepSeek

685.0B

Best score:0.8 (GPQA)

Released:Sep 2025

Price:$0.27/1M tokens

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.