DeepSeek V3.1 Terminus (Thinking)

Thinking-enabled DeepSeek-V3.1-Terminus with improved language consistency, upgraded Code/Search Agents, and stronger stability and reliability versus V3.1. FP8.

Tool Calling
Structured Output

Model weights

Pricing

Auto routing · per 1M tokens

Input: $0.25
Output: $0.70

Compare provider prices

Specifications

Context window: 128K
Max output: 65.5K
Parameters: 671B / 37B; Total / active
Avg output (7d): 289 tokens; Longer than 24% of models

Benchmarks

Artificial Analysis

LMArena

Sourced from Artificial Analysis.

Intelligence Index

14.8

Better than 55% of models compared

Coding Index

43.5

Better than 51% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

2.7%

Better than 23% of models compared

AA-Briefcase

Agentic knowledge work (Elo)

454 Elo

Better than 22% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

712 Elo

Better than 33% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

5.4%

Better than 27% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

69.3%

Better than 61% of models compared

Reasoning

HLE

Humanity's Last Exam

16.4%

Better than 63% of models compared

IFBench

Instruction-following benchmark

57.0%

Better than 68% of models compared

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

0.0%

Better than 16% of models compared

SciCode

Python programming for scientific computing

38.0%

Better than 20% of models compared

LiveCodeBench

Contamination-free coding benchmark

79.8%

Better than 92% of models compared

Math

AIME 2025

American Invitational Mathematics Examination 2025

89.7%

Better than 90% of models compared

Knowledge

MMLU-Pro

Professional and academic subject knowledge

85.1%

Better than 91% of models compared

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

79.2%

Better than 66% of models compared

Terminal-Bench Hard (legacy)

Agentic coding and terminal use

30.3%

Better than 72% of models compared

T²-Bench Telecom (legacy)

Conversational AI agents in dual-control scenarios

37.1%

Better than 45% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation