Gemini 3 Flash
Gemini 3 Flash — проприетарная мультимодальная модель рассуждений компании Google DeepMind, построенная на рассуждающей основе Gemini 3 Pro и оптимизированная по задержке и стоимости инференса, с поддержкой настраиваемых уровней глубины рассуждений для выполнения агентных, программных и мультимодальных задач. Модель была представлена 17 декабря 2025 года.
Оценка
Gemini 3 Flash оценивалась по широкому набору тестов (бенчмарков), охватывающих способности к рассуждению, мультимодальные возможности, использование инструментов в агентных сценариях, многоязычную обработку данных и работу с длинным контекстом. Дополнительные тесты, сведения о подходе к оценке, результатах и методологиях опубликованы в методологии оценки Gemini 3 Flash.
Согласно Model Card Google DeepMind, Gemini 3 Flash значительно превосходит Gemini 2.5 Pro в ряде бенчмарков, требующих развитых способностей к рассуждению и мультимодальной обработке данных. Приведённые ниже результаты актуальны по состоянию на декабрь 2025 года.
| Benchmark | Description | Gemini 3 Flash Thinking |
Gemini 3 Pro Thinking |
Gemini 2.5 Flash Thinking |
Gemini 2.5 Pro Thinking |
Claude Sonnet 4.5 Thinking |
GPT-5.2 Extra high |
Grok 4.1 Fast Reasoning | |
|---|---|---|---|---|---|---|---|---|---|
| Humanity's Last Exam | Academic reasoning (full set, text + MM) |
No tools With search and code execution |
33.7% 43.5% |
37.5% 45.8% |
11.0% — |
21.6% — |
13.7% — |
34.5% 45.5% |
17.6% — |
| ARC-AGI-2 | Visual reasoning puzzles | ARC Prize Verified | 33.6% | 31.1% | 2.5% | 4.9% | 13.6% | 52.9% | — |
| GPQA Diamond | Scientific knowledge | No tools | 90.4% | 91.9% | 82.8% | 86.4% | 83.4% | 92.4% | 84.3% |
| AIME 2025 | Mathematics | No tools With code execution |
95.2% 99.7% |
95.0% 100% |
72.0% 75.7% |
88.0% — |
87.0% 100% |
100% — |
91.9% — |
| MMMU-Pro | Multimodal understanding and reasoning |
81.2% | 81.0% | 66.7% | 68.0% | 68.0% | 79.5% | 63.0% | |
| ScreenSpot-Pro | Screen understanding | No tools unless specified |
69.1% | 72.7% | 3.9% | 11.4% | 36.2% | 86.3% with python |
— |
| CharXiv Reasoning | Information synthesis from complex charts |
No tools | 80.3% | 81.4% | 63.7% | 69.6% | 68.5% | 82.1% | — |
| OmniDocBench 1.5 | OCR | Overall Edit Distance, lower is better |
0.121 | 0.115 | 0.154 | 0.145 | 0.145 | 0.143 | — |
| Video-MMMU | Knowledge acquisition from videos |
86.9% | 87.6% | 79.2% | 83.6% | 77.8% | 85.9% | — | |
| LiveCodeBench Pro | Competitive coding problems from Codeforces, ICPC, and IOI |
Elo rating, higher is better |
2316 | 2439 | 1143 | 1775 | 1418 | 2393 | — |
| Terminal-bench 2.0 | Agentic terminal coding | Terminus-2 harness | 47.6% | 54.2% | 16.9% | 32.6% | 42.8% | — | — |
| SWE-bench Verified | Agentic coding | Single attempt | 78.0% | 76.2% | 60.4% | 59.6% | 77.2% | 80.0% | 50.6% |
| τ2-bench | Agentic tool use | 90.2% | 90.7% | 79.5% | 77.8% | 87.2% | — | — | |
| Toolathlon | Long horizon real-world software tasks | 49.4% | 36.4% | 3.7% | 10.5% | 38.9% | 46.3% | — | |
| MCP Atlas | Multi-step workflows using MCP | 57.4% | 54.1% | 3.4% | 8.8% | 43.8% | 60.6% | — | |
| Vending-Bench 2 | Agentic long term coherence | Net worth (mean), higher is better |
$3,635 | $5,478 | $549 | $574 | $3,839 | $3,952 | $1,107 |
| FACTS Benchmark Suite | Factuality benchmark across grounding, parametric, search, and MM |
61.9% | 70.5% | 50.4% | 63.4% | 48.9% | 61.4% | 42.1% | |
| SimpleQA Verified | Parametric knowledge | 68.7% | 72.1% | 28.1% | 54.5% | 29.3% | 38.0% | 19.5% | |
| MMMLU | Multilingual Q&A | 91.8% | 91.8% | 86.6% | 89.5% | 89.1% | 89.6% | 86.8% | |
| Global PIQA | Commonsense reasoning across 100 Languages and Cultures |
92.8% | 93.4% | 90.2% | 91.5% | 90.1% | 91.2% | 85.6% | |
| MRCR v2 (8-needle) | Long context performance | 128K (average) 1M (pointwise) |
67.2% 22.1% |
77.0% 26.3% |
54.3% 21.0% |
58.0% 16.4% |
47.1% not supported |
81.9% not supported |
54.6% 6.1% |