DeepSeek V3.1

Aug 2025
DeepSeek V3.1
Compare in
on
—per task?
$0.25 / $0.95per 1M in / out?
141 tok/sfirst token in 3.2 s?
128Kcontext

Rankingin thinking mode?

Strong at

B
Reasoning#146 of 334Science#128 of 312Knowledge#148 of 318
C
Tool calling#129 of 242Long context#163 of 273Instructions#148 of 245

Alternatives?

ModelScoreCostPer taskWhat changes
CheaperMiniMax M3.1 Flashthinking×3.4-38%$0.22Complete model response must be appendedBetterGPT-6 Lunalow+42%-53%$0.005Function calling only with none · Chat Completions+3OpenMiMo v2.6 Prothinking×3.5+28%$0.13Memory (4-bit): 411 GB → 615 GB+4CheaperClaude Haiku 5.5max×3.1-53%$0.21Remove sampling parameters+5CheaperGLM 5.3 FlashmaxOpen×2.8-44%$0.28Defaults to max · Self-hosted+4

Run it yourself

estimate
Fits on✗ 8 GB (phone, laptop)✗ 16 GB✗ 24 GB GPU✗ 48 GB GPU✗ 80 GB GPU
LicenceMITfree for commercial use

History

5 versions
0510152025Jan 25Apr 25Jul 25Oct 25Release dateHumanity's Last ExamV3 · DecV3 · MarV3.1V3.2
0510152025Jan 25Jul 25Release dateHumanity's Last ExamV3 · DecV3 · MarV3.1V3.2

Benchmarks

all categories
Chat & Knowledge
Humanity's Last Exam14.3#147
Arena Text · overall1,416#82
AA Omniscience Index-29.6#149
CritPt (physics)2#105
AA Intelligence Index 4.3.2≈13.5#158
AA Omniscience · Accuracy29#85
AA Omniscience · Hallucination Rate82.5#184
AA-LCR (long context) 1.156.7#164
GPQA Diamond77.9#129
AIME 2025 · AA archived89.7#16
Coding
Terminal-Bench Hard (version unstated)25#73
LiveCodeBench · AA archived78.4#20
Agents & Tool Use
τ²-bench · AA37.4#131
Writing & Language
Arena Text · creative writing1,401#68
IFBench (instructions)41.5#149

FAQ

How much does DeepSeek V3.1 cost?›

$0.25 per 1M input tokens and $0.95 per 1M output tokens (cheapest host: DeepInfra).

Is DeepSeek V3.1 free to use?›

There is no free API; the cheapest paid price is $0.25 per 1M input tokens. The weights are open, so you can run it yourself.

How good is DeepSeek V3.1, and what is it best at?›

#146 of 334 in Chat & Knowledge, #73 of 239 in Coding, #129 of 242 in Agents & Tool Use, #68 of 195 in Writing & Language.

Is DeepSeek V3.1 better than DeepSeek V3?›

On Humanity's Last Exam, DeepSeek V3.1 scores 14.3 vs 2.9 — higher.

How can I use DeepSeek V3.1?›

Through the API of AWS Bedrock, CoreWeave, DeepInfra, Featherless, SambaNova and 1 more.

What is the context window of DeepSeek V3.1?›

128K tokens, with up to 33K output tokens.

When was DeepSeek V3.1 released? Is there a newer version?›

Released 2025-08-21. It has been superseded by DeepSeek V3.1 Terminus.

Is DeepSeek V3.1 open source?›

Yes, the weights are open (MIT licence).

Can I run DeepSeek V3.1 locally?›

The weights are open and can be self-hosted; hardware needs are not confirmed yet.

Can DeepSeek V3.1 read images?›

No, it accepts text only.

What is DeepSeek V3.1?›

DeepSeek V3.1 is an open-weights language model from DeepSeek, released Aug 2025, the successor to DeepSeek V3.

What are the best alternatives to DeepSeek V3.1?›

MiniMax M3.1 Flash (cheaper), GPT-6 Luna (better), MiMo v2.6 Pro (open weights).