Skip to content
VECTOR WIREAI INTELLIGENCE
⌘K
UTC
Models739Deals30d267Policy30d64, 15 pendingSources1,010

DeepSeek-V3.2

DeepSeekReleased Dec 1, 2025Deprecated

Score basis

DeepSeek-V3.2 is capable in long context; and behind the leaders in reasoning, instruction following, coding, and agentic tasks. Too few results yet to rate safety, math, multimodal tasks, multilingual tasks, or factuality.

Capability profile

Bars show the model's median result as a share of the leading model's, per capability. Select a row to see its results.

Capable1 capability
  1. Long Context−21.5%2 of 3
Limited4 capabilities
  1. Reasoning−32.5%4 of 6
  2. Instruction Following−33.7%1 of 3
  3. Coding−35.7%8 of 10
  4. Agentic−42.6%4 of 7
Not rated5 capabilities

Too few results yet to rate Safety, Math, Multimodal, Multilingual or Factuality.

Price

$0.28input$0.42outputper million tokens

From Artificial Analysis · 3 providers tracked · All prices

Cheaper than 71% of 328 priced models · 3:1 input-to-output blend, log scale

Evidence

182results on102benchmarks

  • 26 independently verified
  • 33 aggregator
  • 17 vendor-reported
  • 106 cross-referenced

From 26 sources · latest Oct 7, 2026 · How verification works

API features

Tool callingStructured outputsReasoning

As listed by OpenRouter

Research

33 papers reference DeepSeek-V3.2

DeepSeek-V3.2 benchmark results

182 results on 102 benchmarks, grouped by capability. Each bar is the result as a share of the capability leader's; every result links to its source.

Long Context

CapableFull long context ranking

21.5% behind the leader2 of 3 ranked benchmarks measured

Show 2 more long context resultsHide 2 long context results

Reasoning

LimitedFull reasoning ranking

32.5% behind the leader4 of 6 ranked benchmarks measured

Show 13 more reasoning resultsHide 13 reasoning results

Instruction Following

LimitedFull instruction following ranking

33.7% behind the leader1 of 3 ranked benchmarks measured

Show 3 more instruction following resultsHide 3 instruction following results

Coding

LimitedFull coding ranking

35.7% behind the leader8 of 10 ranked benchmarks measured

Show 19 more coding resultsHide 19 coding results

Agentic

LimitedFull agentic ranking

42.6% behind the leader4 of 7 ranked benchmarks measured

Show 4 more agentic resultsHide 4 agentic results

Math

Not enough dataFull math ranking

0 of 5 ranked benchmarks measured

Show 6 more math resultsHide 6 math results

Factuality

Not enough dataFull factuality ranking

0 of 4 ranked benchmarks measured

Show 3 more factuality resultsHide 3 factuality results

More results

Benchmarks outside the capability baskets. They are not ranked against a leader.

Show 101 more resultsHide 101 results

DeepSeek-V3.2: common questions

Who makes DeepSeek-V3.2?

DeepSeek-V3.2 is made by DeepSeek.

When was DeepSeek-V3.2 released?

DeepSeek-V3.2 was released on Dec 1, 2025, according to Artificial Analysis.

What is DeepSeek-V3.2 good at?

DeepSeek-V3.2 is capable in long context; and behind the leaders in reasoning, instruction following, coding, and agentic tasks. Too few results yet to rate safety, math, multimodal tasks, multilingual tasks, or factuality.

How much does DeepSeek-V3.2 cost?

DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens, according to Artificial Analysis. We track its price at 3 providers. At a mix of three input tokens to one output token, it is cheaper than 71% of the 328 priced models we track.

How many benchmarks has DeepSeek-V3.2 been tested on?

We track 182 results for DeepSeek-V3.2 on 102 benchmarks from 26 sources, 26 of them independently verified. The latest was recorded on Oct 7, 2026.

Which API features does DeepSeek-V3.2 support?

OpenRouter lists tool calling, structured outputs, and reasoning for DeepSeek-V3.2.

About this record

Where DeepSeek-V3.2's numbers come from, and every name it appears under.

Tracked since
Apr 25, 2026
Newest source mention
Aug 23, 2026

Where the results come from

Verification: 182 scores · 26 independently verified · 33 aggregator-attributed · 106 vendor cross-reference · 17 vendor-reported. How these tiers are assigned

From 26 sources on 10 sites. Hugging Face supplies 110 of them; the 26 independently verified results come from 8 sites. Bars are coloured by trust tier.

  • huggingface.co110
  • artificialanalysis.ai33
  • api.llm-stats.com17
  • matharena.ai7
  • raw.githubusercontent.com4
  • swebench.com4
  • arcprize.org2
  • datasets-server.huggingface.co2
  • epoch.ai2
  • labs.scale.com1

Also known as

deepseek v3.2 (high reasoning)deepseek v3.2 (non-reasoning)deepseek v3.2 (reasoning)deepseek v3.2 reasonerdeepseek-v3-2deepseek-v3-2 (Non-Reasoning)deepseek-v3-2 (reasoning)deepseek-v3-2-reasoningdeepseek-v3.2 (think)deepseek-v3.2 (thinking; fireworks)deepseek-v3.2 (thinking)deepseek-v3.2 thinkingdeepseek-v3.2-20251201deepseek-v3.2-exp (think)deepseek-v3p2deepseekv3.2and 1 more