Noometry

DeepSeek, open weights

DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp by DeepSeek ranks 78th of 354 ranked models on the Noometry Index as of October 2026, with a score of 44.3. Its strongest category is long context, where it ranks 16th. API pricing starts at $0.26 per million input tokens and $0.38 per million output tokens, with a 164K-token context window.

Last verified

Noometry rank
#78 of 354
Index score
44.3
Evidence
Confirmed 49 results
Provider
DeepSeek
Released
September 29, 2025
Weights
Open weights
Reasoning
Yes
Context window
164K
Max output
66K
Input price
$0.26 / M
Output price
$0.38 / M
Blended price
$0.29 / M
Output speed
16 tokens/s Kagi
Value
#49 of 219
Knowledge cutoff
December 2024
Input
text
Hugging Face
deepseek-ai/DeepSeek-V3.2-Exp

Category scores

Each category score combines every public result we have in that category.

  1. Coding 46.5
  2. Agentic & Tool Use 32.7
  3. Reasoning 22.1
  4. Math 41.7
  5. Knowledge 51.7
  6. Multilingual 52.2
  7. Instruction Following 74.5
  8. Long Context 47.6
  9. Writing & Preference 62.4
CategoryScoreRankResults
Coding46.5#657
Agentic & Tool Use32.7#594
Reasoning22.1#20810
Math41.7#874
Knowledge51.7#663
Multilingual52.2#901
Instruction Following74.5#931
Long Context47.6#164
Writing & Preference62.4#774

Strengths and weaknesses

Categories where DeepSeek-V3.2-Exp places highest and lowest among the models ranked in each, with its score against that category's median.

Strongest categories

CategoryScorevs medianRank
Long Context47.6+6.6#16 of 296, top 6%
Coding46.5+7.8#65 of 340, top 20%
Knowledge51.7+14.3#66 of 314, top 22%

Weakest categories

CategoryScorevs medianRank
Reasoning22.1−1.5#208 of 350, top 60%
Agentic & Tool Use32.7+2.4#59 of 154, top 39%
Instruction Following74.5+3.2#93 of 305, top 31%

Closest competitors

The models ranked just above and below DeepSeek-V3.2-Exp. When scores are this close, price and speed are often the better way to choose.

ModelRankScoreBlended $/MSpeed
MiMo-V2.5-Pro#7445.2$0.54—Compare
Gemini 2.5 Pro#7545.0$3.445Compare
GPT-5.4 mini#7645.0$1.6910Compare
Amazon Nova Experimental Chat 26 02 10#7744.5——Compare
Hy3#7944.2$0.14—Compare
Inkling#8044.1$2.57—Compare
Claude Sonnet 4.5#8144.1$685Compare
Chatgpt 4o Latest 20250326#8243.8—21Compare

Sponsored placements are available on pages like this one. Advertise on Noometry

Benchmark results

Every published result we track, with its source. Bold rows are the ones used for ranking; where several exist we prefer independent runs over self-reported numbers.

Coding

BenchmarkScorePositionSettingSourceDate
SWE-bench Verified (bash only)70%#11 of 39, top 29%highSWE-bench2026-02-17
Aider Polyglot70.2%Epoch AI
Aider Polyglot74.2%Epoch AI
Aider Polyglot70.2%Epoch AI
Aider Polyglot74.2%#6 of 44, top 14%thinkingEpoch AI
LMArena WebDev1272LMArena2026-10-08
LMArena WebDev1362#83 of 113, top 74%thinkingLMArena2026-10-08
SWE-bench Multilingual59%#12 of 13, top 93%SWE-bench2026-02-13
SciCode38.9%#85 of 121, top 71%thinkingEpoch AI
WeirdML39.5%#79 of 119, top 67%thinkingEpoch AI
LMArena Coding1454#80 of 294, top 28%thinkingLMArena2026-10-08
LMArena Coding1439thinkingLMArena2026-10-08

Agentic & Tool Use

BenchmarkScorePositionSettingSourceDate
Terminal-Bench39.6%#24 of 41, top 59%Epoch AI
APEX-Agents21.3%#48 of 49, top 98%Epoch AI
Berkeley Function Calling Leaderboard56.7%#11 of 49, top 23%prompt + thinkingBerkeley Function Calling Leaderboard
TheAgentCompany42.9%Best of 14Epoch AI
Vending-Bench 21,034#47 of 60, top 79%Epoch AI

Reasoning

BenchmarkScorePositionSettingSourceDate
ARC-AGI-24%#62 of 83, top 75%Epoch AI
Kagi LLM Benchmark52.2%#60 of 99, top 61%Kagi LLM Benchmark
NYT Connections (extended)36.7%#70 of 91, top 77%Lech Mazur benchmarks
ARC-AGI-157%#53 of 83, top 64%Epoch AI
CritPt2.9%#68 of 134, top 51%thinkingEpoch AI
Chess Puzzles14%#69 of 129, top 54%Epoch AI2025-12-16
Chess Puzzles1%Epoch AI2026-07-16
Thematic Generalization65%#9 of 23, top 40%Lech Mazur benchmarks
LMArena Hard Prompts1434#86 of 297, top 29%LMArena2026-10-08
LMArena Hard Prompts1429LMArena2026-10-08
DTBench62.7%Epoch AI
DTBench85.6%Epoch AI
DTBench87.7%#47 of 151, top 32%thinkingEpoch AI
LMCA28.8%Epoch AI
LMCA15.2%Epoch AI
LMCA29.1%#78 of 125, top 63%thinkingEpoch AI
Epoch Capabilities Index145Epoch AI2025-09-29
Epoch Capabilities Index146.27#74 of 213, top 35%Epoch AI2025-12-01

Math

BenchmarkScorePositionSettingSourceDate
MathArena Final-Answer Competitions57.7%#24 of 29, top 83%thinkMathArena
OTIS Mock AIME 2024-202548.9%Epoch AI2026-07-16
OTIS Mock AIME 2024-202587.8%#59 of 173, top 35%Epoch AI2025-12-16
ProofBench8%#66 of 77, top 86%Epoch AI
LMArena Math1435#78 of 285, top 28%LMArena2026-10-08
LMArena Math1423thinkingLMArena2026-10-08
FrontierMath (Feb 2025 set)22.1%#27 of 68, top 40%Epoch AI2025-12-22
FrontierMath Tier 4 (v1)2.1%#37 of 55, top 68%Epoch AI2025-12-16

Knowledge

BenchmarkScorePositionSettingSourceDate
GPQA Diamond83.4%#69 of 186, top 38%Epoch AI2025-12-16
GPQA Diamond71.2%Epoch AI2026-07-16
Vectara Hallucination Rate (lower is better)6.3%Vectara Hallucination Leaderboard
Vectara Hallucination Rate (lower is better)5.3%#13 of 96, top 14%Vectara Hallucination Leaderboard
LMArena Expert1436#89 of 273, top 33%LMArena2026-10-08
LMArena Expert1422thinkingLMArena2026-10-08

Multilingual

BenchmarkScorePositionSettingSourceDate
LMArena Non-English1409#90 of 297, top 31%LMArena2026-10-08
LMArena Non-English1404thinkingLMArena2026-10-08
LMArena Chinese1456LMArena2026-10-08
LMArena Chinese1461#89 of 285, top 32%LMArena2026-10-08
LMArena French1433#87 of 223, top 40%LMArena2026-10-08
LMArena French1429LMArena2026-10-08
LMArena German1440#57 of 231, top 25%LMArena2026-10-08
LMArena German1413thinkingLMArena2026-10-08
LMArena Japanese1331LMArena2026-10-08
LMArena Japanese1374#85 of 211, top 41%thinkingLMArena2026-10-08
LMArena Korean1371#85 of 213, top 40%LMArena2026-10-08
LMArena Korean1370LMArena2026-10-08
LMArena Russian1424#75 of 283, top 27%LMArena2026-10-08
LMArena Russian1411thinkingLMArena2026-10-08
LMArena Spanish1440#66 of 226, top 30%LMArena2026-10-08
LMArena Spanish1423LMArena2026-10-08

Instruction Following

BenchmarkScorePositionSettingSourceDate
LMArena Instruction Following1413#83 of 298, top 28%LMArena2026-10-08
LMArena Instruction Following1401LMArena2026-10-08

Long Context

BenchmarkScorePositionSettingSourceDate
Fiction.LiveBench52.8%Epoch AI
Fiction.LiveBench83.3%#9 of 47, top 20%highEpoch AI
CL-bench12.4%Epoch AI
CL-bench13.2%#18 of 19, top 95%thinkingEpoch AI
CL-bench Life7.4%Epoch AI
CL-bench Life9.5%#11 of 13, top 85%thinkingEpoch AI
LMArena Longer Query1428#79 of 291, top 28%LMArena2026-10-08
LMArena Longer Query1418LMArena2026-10-08

Writing & Preference

BenchmarkScorePositionSettingSourceDate
LMArena Text1425LMArena2026-10-08
LMArena Text1425#87 of 297, top 30%thinkingLMArena2026-10-08
LMArena Creative Writing1403#71 of 295, top 25%LMArena2026-10-08
LMArena Creative Writing1400LMArena2026-10-08
EQ-Bench Creative Writing1515#53 of 115, top 47%EQ-Bench
LMArena Multi-Turn1427#88 of 295, top 30%LMArena2026-10-08
LMArena Multi-Turn1422LMArena2026-10-08

API pricing by provider

RouteInput $/MOutput $/MCached input $/MChecked
azure$0.58$1.68—2026-10-10
bedrock$0.62$1.85—2026-10-10
deepinfra$0.26$0.38$0.132026-10-10
openrouter$0.27$0.41—2026-10-10
vertex$0.56$1.68$0.0562026-10-10

All DeepSeek API prices → Estimate your cost →

Compare DeepSeek-V3.2-Exp

Other DeepSeek models

Frequently asked questions

How good is DeepSeek-V3.2-Exp?

DeepSeek-V3.2-Exp by DeepSeek ranks 78th of 354 ranked models on the Noometry Index as of October 2026, with a score of 44.3. Its strongest category is long context, where it ranks 16th. API pricing starts at $0.26 per million input tokens and $0.38 per million output tokens, with a 164K-token context window.

How much does DeepSeek-V3.2-Exp cost?

DeepSeek-V3.2-Exp costs $0.26 per million input tokens and $0.38 per million output tokens on deepinfra, with cached input at $0.13.

What is DeepSeek-V3.2-Exp's context window?

DeepSeek-V3.2-Exp accepts up to 164K tokens of input and can write up to 66K tokens in one response.

Is DeepSeek-V3.2-Exp open source?

Yes. DeepSeek-V3.2-Exp's weights are downloadable from Hugging Face (deepseek-ai/DeepSeek-V3.2-Exp); check the license for commercial terms.

How fast is DeepSeek-V3.2-Exp?

DeepSeek-V3.2-Exp generated about 16 output tokens per second in the Kagi LLM Benchmark's timed runs. Speed varies by provider, load and reasoning effort.

What are DeepSeek-V3.2-Exp's strengths and weaknesses?

Relative to other ranked models, DeepSeek-V3.2-Exp places best in long context, coding, knowledge and lowest in reasoning, agentic & tool use, instruction following.

What is DeepSeek-V3.2-Exp best at?

Its best category is long context, where it ranks 16th on Noometry.