Model comparison
DeepSeek-V2.5 (Sep 2024) vs DeepSeek-V3
DeepSeek-V3 is the stronger model overall, scoring 39.5 to 37.6 on the Noometry Index.
Last verified . 22 shared benchmarks.
DeepSeekSummary
- They share 22 benchmarks with published results for both. DeepSeek-V2.5 (Sep 2024) scores higher in 3 categories and DeepSeek-V3 in 5 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in coding, where DeepSeek-V3 leads 42.3 to 31.7.
- The biggest single-benchmark swing is Aider Polyglot: 17.8% for DeepSeek-V2.5 (Sep 2024) and 55.1% for DeepSeek-V3.
Side by side
| DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 | |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Noometry Index | 37.6 | 39.5 |
| Released | 2024-09-06 | 2024-12-26 |
| Weights | Open | Open |
| Context window | — | 164K |
| Max output | — | 164K |
| Input $ / M tokens | — | $0.24 |
| Output $ / M tokens | — | $0.90 |
| Results tracked | 22 | 60 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-V3 leads
DeepSeek-V2.5 (Sep 2024): 31.7 (#281), DeepSeek-V3: 42.3 (#106)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| Aider Polyglot | 17.8% | 55.1% |
| BigCodeBench Instruct | 48.6% | 50% |
| LMArena Coding | 1309 | 1368 |
| BigCodeBench Complete | 53.2% | 62.2% |
| HumanEval+ | 83.5% | 86.6% |
| MBPP+ | 74.1% | 73% |
| SciCode | — | 35.8% |
| WeirdML | — | 36.1% |
| LiveBench Coding | — | 70.9% |
Agentic & Tool Use Not comparable
DeepSeek-V2.5 (Sep 2024): —, DeepSeek-V3: —
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| METR Time Horizons | — | 49.6% |
Reasoning DeepSeek-V2.5 (Sep 2024) leads
DeepSeek-V2.5 (Sep 2024): 25.6 (#145), DeepSeek-V3: 20.5 (#236)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Hard Prompts | 1289 | 1365 |
| SimpleBench | — | 27.2% |
| Kagi LLM Benchmark | — | 52.3% |
| CritPt | — | 0% |
| LiveBench Reasoning | — | 65.8% |
| DTBench | — | 64.8% |
| LiveBench Data Analysis | — | 60.9% |
| LMCA | — | 15.5% |
| BIG-Bench Hard | — | 87.5% |
| Epoch Capabilities Index | — | 135.94 |
| ForecastBench | — | 59.1 |
| HellaSwag | — | 88.9% |
| LiveBench | — | 66.9% |
| PIQA | — | 84.7% |
| WinoGrande | — | 85.2% |
Math DeepSeek-V2.5 (Sep 2024) leads
DeepSeek-V2.5 (Sep 2024): 35.9 (#177), DeepSeek-V3: 32.1 (#219)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Math | 1288 | 1373 |
| OTIS Mock AIME 2024-2025 | — | 37.8% |
| Omni-MATH | — | 40.3% |
| LiveBench Math | — | 73.5% |
| MATH Level 5 | — | 75.5% |
| FrontierMath (Feb 2025 set) | — | 1.7% |
Knowledge DeepSeek-V3 leads
DeepSeek-V2.5 (Sep 2024): 34.8 (#193), DeepSeek-V3: 37.5 (#155)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Expert | 1266 | 1351 |
| GPQA Diamond | — | 67.6% |
| MMLU-Pro | — | 72.3% |
| Confabulations | — | 26.1% |
| Vectara Hallucination Rate | — | 6.1% |
| GPQA (HELM) | — | 53.8% |
| ARC (AI2) Challenge | — | 95.3% |
| MMLU | — | 87.2% |
| TriviaQA | — | 82.9% |
Multilingual DeepSeek-V3 leads
DeepSeek-V2.5 (Sep 2024): 42.5 (#193), DeepSeek-V3: 48.5 (#143)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Non-English | 1273 | 1358 |
| LMArena Chinese | 1318 | 1391 |
| LMArena French | 1289 | 1385 |
| LMArena German | 1258 | 1374 |
| LMArena Japanese | 1228 | 1333 |
| LMArena Korean | 1209 | 1319 |
| LMArena Russian | 1289 | 1373 |
| LMArena Spanish | 1248 | 1358 |
Instruction Following DeepSeek-V3 leads
DeepSeek-V2.5 (Sep 2024): 67.5 (#194), DeepSeek-V3: 72.8 (#130)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Instruction Following | 1280 | 1345 |
| LiveBench Instruction Following | — | 81.5% |
| IFEval | — | 83.2% |
Long Context DeepSeek-V2.5 (Sep 2024) leads
DeepSeek-V2.5 (Sep 2024): 39.5 (#174), DeepSeek-V3: 34.0 (#253)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Longer Query | 1301 | 1352 |
| Fiction.LiveBench | — | 50% |
Writing & Preference DeepSeek-V3 leads
DeepSeek-V2.5 (Sep 2024): 49.8 (#187), DeepSeek-V3: 57.4 (#130)
| Benchmark | DeepSeek-V2.5 (Sep 2024) | DeepSeek-V3 |
|---|---|---|
| LMArena Text | 1294 | 1375 |
| LMArena Creative Writing | 1285 | 1364 |
| LMArena Multi-Turn | 1297 | 1389 |
| Short-Story Creative Writing | — | 77% |
| EQ-Bench Creative Writing | — | 1472 |
| WildBench | — | 83% |
| LiveBench Language | — | 49.1% |
Frequently asked questions
Is DeepSeek-V2.5 (Sep 2024) better than DeepSeek-V3?
DeepSeek-V3 is the stronger model overall, scoring 39.5 to 37.6 on the Noometry Index.
Is DeepSeek-V2.5 (Sep 2024) or DeepSeek-V3 better for coding?
DeepSeek-V3 scores higher on coding benchmarks: 42.3 versus 31.7 in the Noometry coding category.
How many benchmarks do DeepSeek-V2.5 (Sep 2024) and DeepSeek-V3 share?
22 benchmarks have published results for both models. DeepSeek-V2.5 (Sep 2024) has 22 scored results on Noometry and DeepSeek-V3 has 60.