Model comparison
DeepSeek-V3.1 vs DeepSeek-V3.2-Exp
DeepSeek-V3.2-Exp is the stronger model overall, scoring 44.3 to 42.8 on the Noometry Index.
Last verified . 25 shared benchmarks.
DeepSeekSummary
- They share 25 benchmarks with published results for both. DeepSeek-V3.1 scores higher in 1 category and DeepSeek-V3.2-Exp in 7 categories; 6 gaps are clear of the uncertainty.
- The widest gap is in long context, where DeepSeek-V3.2-Exp leads 47.6 to 36.3.
- The biggest single-benchmark swing is Fiction.LiveBench: 52.8% for DeepSeek-V3.1 and 83.3% for DeepSeek-V3.2-Exp.
- DeepSeek-V3.2-Exp is cheaper at $0.26 / $0.38 per million input/output tokens, against $0.25 / $0.95 for DeepSeek-V3.1.
Side by side
| DeepSeek-V3.1 | DeepSeek-V3.2-Exp | |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Noometry Index | 42.8 | 44.3 |
| Released | 2025-08-21 | 2025-09-29 |
| Weights | Open | Open |
| Context window | 164K | 164K |
| Max output | 8K | 66K |
| Input $ / M tokens | $0.25 | $0.26 |
| Output $ / M tokens | $0.95 | $0.38 |
| Results tracked | 27 | 49 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-V3.2-Exp leads
DeepSeek-V3.1: 40.3 (#144), DeepSeek-V3.2-Exp: 46.5 (#65)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| WeirdML | 38.4% | 39.5% |
| LMArena Coding | 1417 | 1454 |
| SWE-bench Verified (bash only) | — | 70% |
| Aider Polyglot | — | 74.2% |
| LMArena WebDev | — | 1362 |
| SWE-bench Multilingual | — | 59% |
| SciCode | — | 38.9% |
Agentic & Tool Use Not comparable
DeepSeek-V3.1: —, DeepSeek-V3.2-Exp: 32.7 (#59)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| Terminal-Bench | — | 39.6% |
| APEX-Agents | — | 21.3% |
| Berkeley Function Calling Leaderboard | — | 56.7% |
| TheAgentCompany | — | 42.9% |
| Vending-Bench 2 | — | 1,034 |
Reasoning DeepSeek-V3.1 leads
DeepSeek-V3.1: 27.9 (#110), DeepSeek-V3.2-Exp: 22.1 (#208)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| Kagi LLM Benchmark | 53.2% | 52.2% |
| LMArena Hard Prompts | 1417 | 1434 |
| DTBench | 82.7% | 87.7% |
| LMCA | 24.3% | 29.1% |
| Epoch Capabilities Index | 139.92 | 146.27 |
| ARC-AGI-2 | — | 4% |
| SimpleBench | 40% | — |
| NYT Connections (extended) | — | 36.7% |
| ARC-AGI-1 | — | 57% |
| CritPt | — | 2.9% |
| Chess Puzzles | — | 14% |
| Thematic Generalization | — | 65% |
| ForecastBench | 58 | — |
Math DeepSeek-V3.2-Exp leads
DeepSeek-V3.1: 38.9 (#122), DeepSeek-V3.2-Exp: 41.7 (#87)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| LMArena Math | 1420 | 1435 |
| MathArena Final-Answer Competitions | — | 57.7% |
| OTIS Mock AIME 2024-2025 | — | 87.8% |
| ProofBench | — | 8% |
| FrontierMath (Feb 2025 set) | — | 22.1% |
| FrontierMath Tier 4 (v1) | — | 2.1% |
Knowledge DeepSeek-V3.2-Exp leads
DeepSeek-V3.1: 43.7 (#90), DeepSeek-V3.2-Exp: 51.7 (#66)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| Vectara Hallucination Rate | 5.5% | 5.3% |
| LMArena Expert | 1405 | 1436 |
| GPQA Diamond | — | 83.4% |
Multilingual Too close to call
DeepSeek-V3.1: 51.6 (#106), DeepSeek-V3.2-Exp: 52.2 (#90)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| LMArena Non-English | 1400 | 1409 |
| LMArena Chinese | 1469 | 1461 |
| LMArena French | 1447 | 1433 |
| LMArena German | 1411 | 1440 |
| LMArena Japanese | 1378 | 1374 |
| LMArena Korean | 1337 | 1371 |
| LMArena Russian | 1405 | 1424 |
| LMArena Spanish | 1431 | 1440 |
Instruction Following Too close to call
DeepSeek-V3.1: 73.9 (#110), DeepSeek-V3.2-Exp: 74.5 (#93)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| LMArena Instruction Following | 1400 | 1413 |
Long Context DeepSeek-V3.2-Exp leads
DeepSeek-V3.1: 36.3 (#232), DeepSeek-V3.2-Exp: 47.6 (#16)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| Fiction.LiveBench | 52.8% | 83.3% |
| LMArena Longer Query | 1422 | 1428 |
| CL-bench | — | 13.2% |
| CL-bench Life | — | 9.5% |
Writing & Preference DeepSeek-V3.2-Exp leads
DeepSeek-V3.1: 60.3 (#98), DeepSeek-V3.2-Exp: 62.4 (#77)
| Benchmark | DeepSeek-V3.1 | DeepSeek-V3.2-Exp |
|---|---|---|
| LMArena Text | 1420 | 1425 |
| LMArena Creative Writing | 1401 | 1403 |
| EQ-Bench Creative Writing | 1436 | 1515 |
| LMArena Multi-Turn | 1408 | 1427 |
Frequently asked questions
Is DeepSeek-V3.1 better than DeepSeek-V3.2-Exp?
DeepSeek-V3.2-Exp is the stronger model overall, scoring 44.3 to 42.8 on the Noometry Index.
Which is cheaper, DeepSeek-V3.1 or DeepSeek-V3.2-Exp?
DeepSeek-V3.2-Exp is cheaper. It lists at $0.26 per million input tokens and $0.38 per million output tokens; DeepSeek-V3.1 lists at $0.25 and $0.95.
Is DeepSeek-V3.1 or DeepSeek-V3.2-Exp better for coding?
DeepSeek-V3.2-Exp scores higher on coding benchmarks: 46.5 versus 40.3 in the Noometry coding category.
Which has the bigger context window?
Both accept 164K tokens.
How many benchmarks do DeepSeek-V3.1 and DeepSeek-V3.2-Exp share?
25 benchmarks have published results for both models. DeepSeek-V3.1 has 27 scored results on Noometry and DeepSeek-V3.2-Exp has 49.