DeepSeekChina· released Dec 2025· open weights

DeepSeek V3.2

Name: DeepSeek V3.2
Author: DeepSeek

Provisional: 15 comparable tests so far. Ranked once it reaches 6 across 3 counting categories.

DeepSeek, China. One of 4 models from this lab on the Index.

Third-party publishedLow confidence (thin coverage)

Snapshot September 8, 2026

OpenCharts Index

48.3

Provisional: not enough comparable tests or counting categories to rank yet · 15 comparable tests · 2 counting categories

Drop any single test and the Index lands between 43.5 and 53.3.

Needs more results (2 of 8 so far)

Averaging 25.5 on the 2 it has; the score opens at 4 of 8.

Distance to today's frontier on the hardest tests. Not a measure of intelligence.

What it's known for

DeepSeek V3.2, in 15 results.

Strongest in coding (53.1) and reasoning (43.5). It trails the frontier most in reasoning, 56.5% behind on average. No comparable results yet in knowledge, multimodal and long context. Between 50% and 60% behind the best published results, on average.

Strongest category

Coding

Writing, fixing and shipping real code, judged by tests or users.

53.1

Aider Polyglot74.2% · 84% of best
SciCode38.9% · 63% of best
WeirdML46.7% · 50% of best
Code Arena (WebDev)1,360 · 15% win rate vs leader

Mean of 4 of the category's 7 comparable tests. Counts toward the Index.

Best single result

Mock AIME 2024–2025

Solve competition mathematics problems that have a single exact integer answer.

87.8%

87.8% of the best (100.0%) · #47 of 78.

Olympiad-style problems need long, careful chains of reasoning, and there is no partial credit.

Where it trails

ARC-AGI-2

Infer the hidden rule from a few input and output grid pairs, then apply it to a new grid.

4.0%

95.8% behind the best published result (95.0%) · #49 of 51.

Novel visual puzzles resist memorization, so they measure fluid intelligence rather than recall.

Not measured yet

No comparable result yet in these 3 categories. A blank is a blank, never a zero, and it does not lower the Index.

Knowledge
Multimodal
Long context

Covered, but not averaged into the Index:

Math2 of 5 tests · needs 3 · mean 47.9
Agentic3 of 8 tests · needs 4 · mean 28.3
Human preferencesingle-test category, shown beside the Index · 76.8

Every score, with its source

See it at work

What DeepSeek V3.2 was asked to do, and how it did.

Every test it has taken, by category, with the kind of task it faced, who produced the number and where the result landed against the best published one. The examples are original and illustrative, never items from the datasets themselves.

Human preference · mean 76.8 across 1 comparable test · a single-test category, shown beside the Index and never averaged into it

Text ArenaLeaderboard

Answer the same prompt as a rival model; real users vote blind on which answer they prefer.

1,425

38% expected win rate against the board leader (1,507)
#71 of 92

board 2026-09-02

Source: Arena

Every result is read against the best published one on its test: percent scores as a share of the best above chance, Arena ratings as the expected win rate against the board leader, open-ended values on a log scale. Sitting a harder exam never lowers a model. Every number links to the publisher that produced it.

Head to head

DeepSeek V3.2 against whoever you pick.

Choose a rival. Every counted test both have taken appears side by side as a share of the best published result on that test, with the category means above. A win is a higher share; a gap under 2 points is a tie, because one item on a 45-problem exam is 2.2 points and publishers' own standard errors are of that order.

DeepSeek V3.2

provisional·Index 48.3D

0 : 10

wins · 0 ties · 10 shared

GPT-6 Astra

#1·Index 99.0A+

GPT-6 Astra comes out ahead on 10 of the 10 tests both have taken.

Category means

Reasoning
43.5
99.6
Knowledge
—
100.0
Coding
53.1
97.7
Math
47.9
99.8
Agentic
28.3
98.5
Human preference
76.8
—

Shared tests · biggest gaps first

ARC-AGI-2Reasoning
4.0%
95.0%
−95.8
ProofBenchMath
8.0%
99.0%
−91.0
CritPtReasoning
2.9%
31.7%
−89.4
Code Arena (WebDev)Coding
1,360
1,797
−85.0
APEX-AgentsAgentic
7.0%
46.7%
−83.7
WeirdMLCoding
46.7%
92.9%
−49.7
ARC-AGI-1Reasoning
57.0%
98.5%
−42.1
SciCodeCoding
38.9%
56.5%
−28.3
GPQA DiamondReasoning
83.4%
95.8%
−17.4
Mock AIME 2024–2025Math
87.8%
100.0%
−12.2

Bars are distances to the best published result on each test (percent scores above chance, Arena ratings as win rate against the leader, open-ended values on a log scale). Raw values as the publishers report them.

Open GPT-6 Astra

Scores

Every test, every source

The best published run per benchmark, the raw value as the publisher reports it, who produced the number, how close it comes to the best published result on that test, and where the model sits among every model scored on it. Percent scores are read above chance, Arena ratings as a win rate against the board leader, open-ended values on a log scale. A test with too few models to compare against is shown but not counted.

Category means

Reasoning43.5 · 5/6
Knowledge—
Coding53.1 · 4/7
Math47.9 · 2/5
Agentic28.3 · 3/8
Multimodal—
Human preferencebeside76.8 · 1/1
Long context—

Mean · tests taken of the category's comparable tests. A faded category either rests on fewer than half of them or is a single-test category (marked “beside”), which is shown next to the Index and never averaged into it.

GPQA DiamondEpoch-runReasoning

83.4%

82.6% of the best (95.8%), above chance
25th percentile of 78 · n = 198

Source: Epoch AI — AI Benchmarking Hub

ARC-AGI-2LeaderboardReasoning

4.0%

4.2% of the best (95.0%)
2th percentile of 51 · n = 120

Source: Epoch AI — AI Benchmarking Hub

ARC-AGI-1LeaderboardReasoning

57.0%

57.9% of the best (98.5%)
16th percentile of 51 · n = 100

Source: Epoch AI — AI Benchmarking Hub