DeepSeek V3.2

Dec 2025
Older versionReplace withDeepSeek V4 Pro 0813Newer versionMiniMax M3.1 FlashCheaper, as goodGPT-6 LunaBetterMiMo v2.6 ProOpen weights
DeepSeek V3.2
Compare in
on
—per task?
$0.28 / $0.40per 1M in / outcache $0.028
—output speed?
128Kcontext

Rankingin non-thinking mode?

Retired models are left out of the current rankings.

Alternatives?

ModelScoreCostPer taskWhat changes
CheaperClaude Haiku 5.5max×2.7-35%$0.21Remove sampling parameters+5BetterGPT-6 Lunalow+34%-35%$0.005Function calling only with none · Chat Completions+3OpenMiMo v2.6 Prothinking×2.9+75%$0.13Memory (4-bit): 411 GB → 615 GB+4CheaperMiniMax M3.1 Flashthinking×2.7-15%$0.22Complete model response must be appendedCheaperGLM 5.3 FlashmaxOpen×2.6-23%$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
101520Jan 25Apr 25Jul 25Oct 25Release dateAA Intelligence Index 4.3.2V3 · DecV3 · MarV3.1V3.2
101520Jan 25Jul 25Release dateAA Intelligence Index 4.3.2V3 · DecV3 · MarV3.1V3.2

Benchmarks

all categories
Chat & Knowledge
Humanity's Last Exam11.2
AA Omniscience Index-46.9
CritPt (physics)0.9
AA Intelligence Index 4.3.2≈16
AA Omniscience · Accuracy24
AA Omniscience · Hallucination Rate93.3
AA-LCR (long context) 1.145.7
MMLU-Prolab85
AIME 2025lab93.1
HMMT February 2025lab92.5
HMMT November 2025lab90.2
IMOAnswerBenchlab78.3
HMMT 2025 chart edition unstatedlab90.2
Humanity’s Last Examlab19.8
HMMT 2025lab83.6
GPQA Diamond75.1
AIME 2025 · AA archived59
Coding
Terminal-Bench Hard (version unstated)32.6
Terminal Bench 2.0lab46.4
LiveCodeBenchlab83.3
Aider Polyglotlab74.5
Codeforces Ratinglab2,386
SWE Verifiedlab73.1
SWE-bench Multilinguallab57.9
Terminal-bench version unstatedlab37.7
Codeforces-Div1 ratinglab2,121
LiveCodeBench · AA archived59.3
Agents & Tool Use
BrowseComplab40.1
BrowseComp-zhlab47.9
Tau2-Benchlab80.3
MCP-Universelab45.9
MCP-Marklab38
Tool-Decathlonlab35.2
HLE without toolslab25.1
SimpleQA search agentlab97.1
τ²-bench · AA78.9
Writing & Language
IFBench (instructions)49

FAQ

How much does DeepSeek V3.2 cost?›

$0.28 per 1M input tokens and $0.40 per 1M output tokens; cached input $0.028 per 1M; cheapest via DeepInfra at $0.21 / $0.32.

Is DeepSeek V3.2 free to use?›

There is no free API; the cheapest paid price is $0.21 per 1M input tokens. The weights are open, so you can run it yourself.

How can I use DeepSeek V3.2?›

Through the API of AWS Bedrock, AtlasCloud, DeepInfra, DigitalOcean, Featherless and 6 more.

What is the context window of DeepSeek V3.2?›

128K tokens, with up to 164K output tokens.

When was DeepSeek V3.2 released? Is there a newer version?›

Released 2025-12-01. It has been superseded by DeepSeek V4 Pro 0424.

Is DeepSeek V3.2 open source?›

Yes, the weights are open (MIT licence).

Can I run DeepSeek V3.2 locally?›

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

Can DeepSeek V3.2 read images?›

No, it accepts text only.

What is DeepSeek V3.2?›

DeepSeek V3.2 is an open-weights language model from DeepSeek, released Dec 2025, the successor to DeepSeek V3.1 Terminus.

What are the best alternatives to DeepSeek V3.2?›

Claude Haiku 5.5 (cheaper), GPT-6 Luna (better), MiMo v2.6 Pro (open weights).