DeepSeekReleased Dec 1, 2025
Open weights

deepseek-v3.2

Its best results are in Programmatic Reasoning (#90 of 177) and Academic Reasoning (#111 of 203); weakest in Reliability (#143 of 177). Knowledge Work is estimated, because no results are published yet.

Composite IQ ?
101
#118 of 176 ranked models · ahead of 33%
DeepSeek's #6 model, behind DeepSeek V4.1 Flash 1017 (128)
Effective cost
$1.84/1M
$0.28 in · $0.42 out
Speed
33tok/s
60s typical response
Context
Unverified
Not in publisher docs
Evidence
6of 7 dims
18 of 39 benchmarks

IQ profile

Seven dimensions, averaged equally into Composite IQ
Compare shape with
100 avg human130160● Abstract79#160● Math103#119● Academic118#111● Coding111#90● Computer Use101#117● Reliability101#143● Knowledge Work96#131 · est.
MeasuredEstimated95% range
#1 in dimensionTop-10 median

IQ dimensions

Ranks are among models scored in each dimension. Ranges are model-based estimates, not calibrated confidence intervals.

DimensionIQPosition · 70–160RankEvidencevs Top 10
79measured
#160 / 201
3/5
-47.5
103measured
#119 / 176
4/5
-44
118measured
#111 / 203
5/6
-17.5
111measured
#90 / 177
1/6
-38.5
101measured
#117 / 176
1/5
-42
101measured
#143 / 177
4/6
-33
96estimated
#131 / 176
0/6
-35.5
deepseek-v3.295% range (model-based, uncalibrated)Top-10 median#1 in dimension100 = average human

Benchmark results

18 of 39 Composite IQ benchmarks have a published result for deepseek-v3.2. Benchmarks without a result stay listed so gaps are visible.

Abstract Reasoning IQ 79 · measuredBest: #53 of 68 on ARC-AGI-13/5
BenchmarkScorePercentileRankLeader · source
ARC-AGI-1
Full source leaderboard reconciliation · Oct 23
57.0%
#53 / 68
Leader fable-5 · 98.5%
Full source leaderboard reconciliation · Oct 23
Chess Puzzles
Epoch AI vEpoch 100-position leaderboard (task implementation versions recorded per run) · Oct 15
14.0%
#61 / 76
Leader gpt-6-astra · 72.0%
Epoch AI vEpoch 100-position leaderboard (task implementation versions recorded per run) · Oct 15
ARC-AGI-2
Full source leaderboard reconciliation · Oct 23
4.0%
#58 / 68
Leader gpt-6-astra · 95.0%
Full source leaderboard reconciliation · Oct 23
Not yet reported
ARC-AGI-3Mystery Game Puzzles
Mathematical Reasoning IQ 103 · measuredBest: #36 of 56 on AIME4/5
BenchmarkScorePercentileRankLeader · source
84.6%
#36 / 56
Leader gemini-3.1-pro · 98.1%
MathArena
Full source leaderboard reconciliation · Oct 28
15.5%
#49 / 64
Leader gpt-6.1-sol · 84.0%
Full source leaderboard reconciliation · Oct 28
FrontierMath Tier 1-3
Epoch AI FrontierMath Tiers 1-3 v2 · Sep 22
22.1%
#70 / 84
Leader gpt-6-astra · 93.7%
Epoch AI FrontierMath Tiers 1-3 v2 · Sep 22
FrontierMath Tier 4
Epoch AI FrontierMath Tier 4 v2 · Sep 22
2.1%
#64 / 72
Leader gpt-6.1-sol · 100%
Epoch AI FrontierMath Tier 4 v2 · Sep 22
Not yet reported
ProofBench 1.1
Academic Reasoning IQ 118 · measuredBest: #15 of 45 on MMLU-Pro5/6
BenchmarkScorePercentileRankLeader · source
MMLU-Pro
TIGER-Lab MMLU-Pro leaderboard (missing rows only) · Sep 22
86.2%
#15 / 45
Leader fable-5.1 · 92.4%
TIGER-Lab MMLU-Pro leaderboard (missing rows only) · Sep 22
GPQA Diamond
Artificial Analysis model leaderboard · Sep 10
84.0%
#88 / 147
Leader gpt-6-astra · 96.1%
Artificial Analysis model leaderboard · Sep 10
CritPt
Artificial Analysis model leaderboard · Sep 10
3.0%
#96 / 159
Leader Muse Wonder Full · 32.9%
Artificial Analysis model leaderboard · Sep 10
Humanity's Last Exam
Artificial Analysis model leaderboard · Sep 10
22.0%
#113 / 167
Leader opus-5.5 · 61.4%
Artificial Analysis model leaderboard · Sep 10
SciCode
Artificial Analysis model leaderboard · Sep 10
39.0%
#125 / 167
Leader opus-5.5 · 66.9%
Artificial Analysis model leaderboard · Sep 10
Not yet reported
MMMU-Pro
Programmatic Reasoning IQ 111 · measuredBest: #36 of 61 on LiveCodeBench1/6
BenchmarkScorePercentileRankLeader · source
LiveCodeBench
Vals.ai · Aug 4
80.7%
#36 / 61
Leader Fugu Ultra v1.0 · 93.2%
Vals.ai · Aug 4
Not yet reported
IOI V2Terminal-Bench 4.0ProgramBenchSWE-rebenchMirrorCode
Computer Use IQ 101 · measuredBest: #37 of 43 on BrowseComp1/5
BenchmarkScorePercentileRankLeader · source
BrowseComp
Full source leaderboard reconciliation · Oct 19
51.4%
#37 / 43
Leader Kimi K3.1 Pro · 92.1%
Full source leaderboard reconciliation · Oct 19
Not yet reported
OSWorld-VerifiedMCP AtlasArena.ai Agent ArenaAgents' Last Exam
Reliability IQ 101 · measuredBest: #68 of 96 on IFBench4/6
BenchmarkScorePercentileRankLeader · source
IFBench
Artificial Analysis model leaderboard · Aug 19
61.0%
#68 / 96
Leader minimax-m3 · 82.9%
Artificial Analysis model leaderboard · Aug 19
AA Omniscience
Artificial Analysis model leaderboard · Sep 10
-23.0
#115 / 158
Leader opus-5.5 · 46.4
Artificial Analysis model leaderboard · Sep 10
BullshitBench v2
Full source leaderboard reconciliation · Oct 23
13.0%
#68 / 83
Leader opus-4.8 · 94.0%
Full source leaderboard reconciliation · Oct 23
SimpleQA Verified
Epoch AI SimpleQA Verified · Sep 22
27.5%
#67 / 80
Leader Gemini 4 Argon · 76.0%
Epoch AI SimpleQA Verified · Sep 22
Not yet reported
AA Long Context Reasoning v1.1FACTS Grounding
Knowledge Work IQ 96 · estimatedNo published results0/6
No published results yet for Knowledge Work. The 96 shown is estimated from deepseek-v3.2's other dimensions and will be replaced as results come in. How estimates work
AA-Briefcase v1.1GDPval-AA v2.1AutomationBench-AAGDP.pdfAA-AnalystAgentEnterpriseOps-Gym-AA

Release & availability

Release details on record for this model.

Released
Developer
DeepSeek
Weights
Open weightsDownloadable and self-hostable
API price
$0.28 in · $0.42 outper 1M tokens · full pricing

About this score

How much of deepseek-v3.2's IQ is measured, and how the rest is estimated.

18/ 39benchmark results
6/ 7dimensions measured
1estimated
3/5Abstract
4/5Math
5/6Academic
1/6Coding
1/5Computer Use
4/6Reliability
0/6Knowledge Work
Bar height = share of benchmarks with resultsEstimatedNo score
Measured results are shown as published

Every score in the benchmark tables comes from its source unchanged. Composite IQ averages all seven dimensions equally.

Estimates fill the gaps

Knowledge Work has no published results yet, so its IQ is estimated from deepseek-v3.2's other dimensions. Missing benchmarks can also draw on its predecessors, Deepseek V3.1 and Deepseek R1; a predecessor is an estimation input, not evidence that this model has the predecessor's measured results.

A missing result is not a zero

Benchmark versions and evaluation harnesses can differ between sources, so small gaps between models may not be meaningful.

Read the scoring methodology →

Compare deepseek-v3.2

Side-by-side pages for the models it is most often weighed against.

deepseek-v3.2Comparison model
Browse all comparisons →

Frequently asked questions

Quick answers about deepseek-v3.2, drawn from the same data as the rest of this page.

How AI IQ scores models →Get this data from the API →
What is the IQ of deepseek-v3.2?
deepseek-v3.2 has an estimated IQ of 101, which ranks #118 out of 176 public models on the AI IQ leaderboard. The score combines seven equally weighted reasoning dimensions using measured benchmark results and conservative estimates for missing coverage.
What is deepseek-v3.2 best at?
Among the dimensions with published benchmark results, deepseek-v3.2 ranks highest in Programmatic Reasoning (IQ 111, #90 of 177) and Academic Reasoning (IQ 118, #111 of 203). Its lowest-ranked measured dimension is Reliability (IQ 101, #143 of 177).
What is the EQ of deepseek-v3.2?
deepseek-v3.2 doesn't have an EQ score yet. EQ (emotional reasoning) is calculated from EQ-Bench 4 and AttuneBench, and neither has published a result for deepseek-v3.2. AI IQ doesn't estimate EQ from other benchmarks, so the score appears once one of those results is published.
Which benchmarks has deepseek-v3.2 been tested on?
deepseek-v3.2 has published results on 18 of the 39 benchmarks behind AI IQ's Composite IQ, including MMLU-Pro, LiveCodeBench, GPQA Diamond, CritPt and AIME.
How much does deepseek-v3.2 cost?
deepseek-v3.2 costs $0.28 per million input tokens and $0.42 per million output tokens. Adjusted for how many tokens it actually consumes on real tasks, its effective cost works out to $1.84 per million input-plus-output tokens.
Is deepseek-v3.2 open source?
Yes — deepseek-v3.2 is an open-weights model: its trained weights are published, so it can be downloaded and self-hosted.
Who developed deepseek-v3.2?
deepseek-v3.2 was developed by DeepSeek, based in China.
When was deepseek-v3.2 released?
deepseek-v3.2 was publicly released on December 1, 2025.
How large is the context window of deepseek-v3.2?
AI IQ hasn't verified an input context limit for deepseek-v3.2 from its developer's documentation, so none is listed rather than guessing.
How fast is deepseek-v3.2?
In independent measurements, deepseek-v3.2 generates about 33 tokens per second, with a typical response time of about 60 seconds.