Model comparison
DeepSeek V4 Flash vs Gemini 3 Pro
Gemini 3 Pro is the stronger model overall, scoring 54.8 to 53.6 on the Noometry Index.
Last verified . 33 shared benchmarks.
DeepSeek
GoogleSummary
- They share 33 benchmarks with published results for both. DeepSeek V4 Flash scores higher in 2 categories and Gemini 3 Pro in 6 categories; 7 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek V4 Flash leads 60.3 to 49.9.
- The biggest single-benchmark swing is ProofBench: 56% for DeepSeek V4 Flash and 20% for Gemini 3 Pro.
- DeepSeek V4 Flash has downloadable open weights; the other is API-only.
Side by side
| DeepSeek V4 Flash | Gemini 3 Pro | |
|---|---|---|
| Provider | DeepSeek | |
| Noometry Index | 53.6 | 54.8 |
| Released | 2026-04-24 | 2025-11-18 |
| Weights | Open | Proprietary |
| Context window | 1M | — |
| Max output | 393K | — |
| Input $ / M tokens | $0.15 | — |
| Output $ / M tokens | $0.60 | — |
| Results tracked | 41 | 67 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Gemini 3 Pro leads
DeepSeek V4 Flash: 47.9 (#59), Gemini 3 Pro: 51.6 (#39)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena WebDev | 1582 | 1440 |
| WeirdML | 63% | 69.9% |
| LMArena Coding | 1457 | 1481 |
| ALE-Bench | 1,306 | 1,177 |
| SWE-bench Verified | — | 72.9% |
| FrontierCode | 18.8% | — |
| SWE-bench Verified (bash only) | — | 74.2% |
| SWE-bench Multilingual | — | 68.7% |
| SciCode | 49.9% | — |
| GSO | — | 18.6% |
| AlgoTune | — | 1.83 |
Agentic & Tool Use Not comparable
DeepSeek V4 Flash: —, Gemini 3 Pro: 40.6 (#23)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| Terminal-Bench | — | 69.4% |
| Berkeley Function Calling Leaderboard | — | 72.5% |
| GDPval | — | 40.3% |
| Remote Labor Index | — | 1.3% |
| τ²-bench Airline | — | 80.5% |
| τ²-bench Banking | — | 18% |
| τ²-bench Retail | — | 75.9% |
| τ²-bench Telecom | — | 91% |
| DeepResearch Bench | — | 46.3% |
| BALROG | — | 58.1% |
| LMArena Search | — | 1207 |
| METR Time Horizons | — | 71% |
| Vending-Bench 2 | — | 5,478 |
Reasoning DeepSeek V4 Flash leads
DeepSeek V4 Flash: 53.7 (#30), Gemini 3 Pro: 52.5 (#31)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| ARC-AGI-2 | 61.4% | 31.1% |
| SimpleBench | 61.1% | 76.4% |
| Kagi LLM Benchmark | 52.2% | 80.1% |
| NYT Connections (extended) | 89.6% | 94.4% |
| ARC-AGI-1 | 89% | 75% |
| CritPt | 16.6% | 6.9% |
| Chess Puzzles | 33% | 31% |
| LMArena Hard Prompts | 1444 | 1480 |
| Epoch Capabilities Index | 154.49 | 152.92 |
| EnigmaEval | — | 18.2% |
| Mystery Game Puzzles | 34% | — |
| DTBench | 90.9% | — |
| LMCA | 41.7% | — |
| ForecastBench | — | 61.2 |
Math DeepSeek V4 Flash leads
DeepSeek V4 Flash: 60.3 (#37), Gemini 3 Pro: 49.9 (#59)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| MathArena Final-Answer Competitions | 76.5% | 67% |
| OTIS Mock AIME 2024-2025 | 94.4% | 91.4% |
| ProofBench | 56% | 20% |
| LMArena Math | 1427 | 1476 |
| FrontierMath (Tiers 1-3) | 57.5% | — |
| FrontierMath Tier 4 | 24.4% | — |
| Omni-MATH | — | 55.5% |
| FrontierMath (Feb 2025 set) | — | 37.6% |
| FrontierMath Tier 4 (v1) | — | 18.8% |
Knowledge Gemini 3 Pro leads
DeepSeek V4 Flash: 55.4 (#48), Gemini 3 Pro: 64.4 (#16)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| GPQA Diamond | 91% | 92.6% |
| LMArena Expert | 1441 | 1475 |
| Humanity's Last Exam | — | 37.5% |
| SimpleQA Verified | 33.6% | — |
| MMLU-Pro | — | 90.3% |
| Vectara Hallucination Rate | — | 13.6% |
| GPQA (HELM) | — | 80.3% |
Multimodal Not comparable
DeepSeek V4 Flash: —, Gemini 3 Pro: 57.6 (#2)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena Vision | — | 1305 |
| GeoBench | — | 84% |
| VPCT | — | 91% |
| LMArena Document | — | 1434 |
Multilingual Gemini 3 Pro leads
DeepSeek V4 Flash: 53.0 (#72), Gemini 3 Pro: 56.9 (#16)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena Non-English | 1420 | 1474 |
| LMArena Chinese | 1468 | 1523 |
| LMArena French | 1439 | 1492 |
| LMArena German | 1418 | 1515 |
| LMArena Japanese | 1406 | 1510 |
| LMArena Korean | 1384 | 1448 |
| LMArena Russian | 1428 | 1493 |
| LMArena Spanish | 1436 | 1470 |
Instruction Following Gemini 3 Pro leads
DeepSeek V4 Flash: 74.9 (#81), Gemini 3 Pro: 76.3 (#45)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena Instruction Following | 1421 | 1458 |
| IFEval | — | 87.7% |
Long Context Too close to call
DeepSeek V4 Flash: 43.8 (#85), Gemini 3 Pro: 44.0 (#79)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena Longer Query | 1434 | 1471 |
| CL-bench | — | 15.8% |
Writing & Preference Gemini 3 Pro leads
DeepSeek V4 Flash: 63.8 (#61), Gemini 3 Pro: 66.4 (#35)
| Benchmark | DeepSeek V4 Flash | Gemini 3 Pro |
|---|---|---|
| LMArena Text | 1432 | 1479 |
| LMArena Creative Writing | 1403 | 1482 |
| EQ-Bench Creative Writing | 1559 | 1525 |
| LMArena Multi-Turn | 1449 | 1484 |
| WildBench | — | 85.9% |
Frequently asked questions
Is DeepSeek V4 Flash better than Gemini 3 Pro?
Gemini 3 Pro is the stronger model overall, scoring 54.8 to 53.6 on the Noometry Index.
Is DeepSeek V4 Flash or Gemini 3 Pro better for coding?
Gemini 3 Pro scores higher on coding benchmarks: 51.6 versus 47.9 in the Noometry coding category.
How many benchmarks do DeepSeek V4 Flash and Gemini 3 Pro share?
33 benchmarks have published results for both models. DeepSeek V4 Flash has 41 scored results on Noometry and Gemini 3 Pro has 67.