---
title: "Gemini 3 Flash"
source: "https://systems-analysis.info/wiki/Gemini_3_Flash"
wiki: "systems-analysis.info/wiki"
article: "Gemini_3_Flash"
language: "ru"
categories:
  - "Категория:Gemini"
  - "Категория:Google"
  - "Категория:Russian"
  - "Категория:Большие языковые модели"
  - "Категория:Семейства LLM"
revision_id: 550
wiki_created_at: 2026-11-10T23:28:46Z
wiki_modified_at: 2026-11-11T00:38:31Z
downloaded_at: 2026-11-11T00:38:42Z
---

# Gemini 3 Flash

**Gemini 3 Flash** — проприетарная [мультимодальная модель](https://systems-analysis.info/wiki/%D0%9C%D1%83%D0%BB%D1%8C%D1%82%D0%B8%D0%BC%D0%BE%D0%B4%D0%B0%D0%BB%D1%8C%D0%BD%D1%8B%D0%B5_LLM "Мультимодальные LLM") рассуждений компании [Google DeepMind](https://systems-analysis.info/wiki/Google_DeepMind "Google DeepMind"), построенная на рассуждающей основе [Gemini 3 Pro](https://systems-analysis.info/wiki/Gemini_3_Pro "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 <a href="https://deepmind.com/models/evals-methodology/gemini-3-flash" class="external text" rel="nofollow">методологии оценки Gemini 3 Flash</a>.

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.

<table class="wikitable" style="text-align:center;">
<colgroup>
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
<col style="width: 10%" />
</colgroup>
<thead>
<tr class="header">
<th>Benchmark</th>
<th>Description</th>
<th></th>
<th style="background: #e6f3ff">Gemini 3 Flash<br />
Thinking</th>
<th><a href="https://systems-analysis.info/wiki/Gemini_3_Pro" title="Gemini 3 Pro">Gemini 3 Pro</a><br />
Thinking</th>
<th><a href="https://systems-analysis.info/wiki/index.php?title=Gemini_2.5_Flash&amp;action=edit&amp;redlink=1" class="new" title="Gemini 2.5 Flash (страница не существует)">Gemini 2.5 Flash</a><br />
Thinking</th>
<th><a href="https://systems-analysis.info/wiki/Gemini_2.5_Pro" title="Gemini 2.5 Pro">Gemini 2.5 Pro</a><br />
Thinking</th>
<th>Claude Sonnet 4.5<br />
Thinking</th>
<th>GPT-5.2<br />
Extra high</th>
<th>Grok 4.1 Fast<br />
Reasoning</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>Humanity's Last Exam</td>
<td>Academic reasoning<br />
(full set, text + MM)</td>
<td>No tools<br />
With search and<br />
code execution</td>
<td style="background: #e6f3ff">33.7%<br />
43.5%</td>
<td><strong>37.5%</strong><br />
<strong>45.8%</strong></td>
<td>11.0%<br />
—</td>
<td>21.6%<br />
—</td>
<td>13.7%<br />
—</td>
<td>34.5%<br />
45.5%</td>
<td>17.6%<br />
—</td>
</tr>
<tr class="even">
<td>ARC-AGI-2</td>
<td>Visual reasoning puzzles</td>
<td>ARC Prize Verified</td>
<td style="background: #e6f3ff">33.6%</td>
<td>31.1%</td>
<td>2.5%</td>
<td>4.9%</td>
<td>13.6%</td>
<td><strong>52.9%</strong></td>
<td>—</td>
</tr>
<tr class="odd">
<td>GPQA Diamond</td>
<td>Scientific knowledge</td>
<td>No tools</td>
<td style="background: #e6f3ff">90.4%</td>
<td>91.9%</td>
<td>82.8%</td>
<td>86.4%</td>
<td>83.4%</td>
<td><strong>92.4%</strong></td>
<td>84.3%</td>
</tr>
<tr class="even">
<td>AIME 2025</td>
<td>Mathematics</td>
<td>No tools<br />
With code execution</td>
<td style="background: #e6f3ff">95.2%<br />
99.7%</td>
<td>95.0%<br />
<strong>100%</strong></td>
<td>72.0%<br />
75.7%</td>
<td>88.0%<br />
—</td>
<td>87.0%<br />
<strong>100%</strong></td>
<td><strong>100%</strong><br />
—</td>
<td>91.9%<br />
—</td>
</tr>
<tr class="odd">
<td>MMMU-Pro</td>
<td>Multimodal understanding<br />
and reasoning</td>
<td></td>
<td style="background: #e6f3ff"><strong>81.2%</strong></td>
<td>81.0%</td>
<td>66.7%</td>
<td>68.0%</td>
<td>68.0%</td>
<td>79.5%</td>
<td>63.0%</td>
</tr>
<tr class="even">
<td>ScreenSpot-Pro</td>
<td>Screen understanding</td>
<td>No tools<br />
unless specified</td>
<td style="background: #e6f3ff">69.1%</td>
<td>72.7%</td>
<td>3.9%</td>
<td>11.4%</td>
<td>36.2%</td>
<td><strong>86.3%</strong><br />
with python</td>
<td>—</td>
</tr>
<tr class="odd">
<td>CharXiv Reasoning</td>
<td>Information synthesis<br />
from complex charts</td>
<td>No tools</td>
<td style="background: #e6f3ff">80.3%</td>
<td>81.4%</td>
<td>63.7%</td>
<td>69.6%</td>
<td>68.5%</td>
<td><strong>82.1%</strong></td>
<td>—</td>
</tr>
<tr class="even">
<td>OmniDocBench 1.5</td>
<td>OCR</td>
<td>Overall Edit Distance,<br />
lower is better</td>
<td style="background: #e6f3ff">0.121</td>
<td><strong>0.115</strong></td>
<td>0.154</td>
<td>0.145</td>
<td>0.145</td>
<td>0.143</td>
<td>—</td>
</tr>
<tr class="odd">
<td>Video-MMMU</td>
<td>Knowledge acquisition<br />
from videos</td>
<td></td>
<td style="background: #e6f3ff">86.9%</td>
<td><strong>87.6%</strong></td>
<td>79.2%</td>
<td>83.6%</td>
<td>77.8%</td>
<td>85.9%</td>
<td>—</td>
</tr>
<tr class="even">
<td>LiveCodeBench Pro</td>
<td>Competitive coding problems<br />
from Codeforces, ICPC, and IOI</td>
<td>Elo rating,<br />
higher is better</td>
<td style="background: #e6f3ff">2316</td>
<td><strong>2439</strong></td>
<td>1143</td>
<td>1775</td>
<td>1418</td>
<td>2393</td>
<td>—</td>
</tr>
<tr class="odd">
<td>Terminal-bench 2.0</td>
<td>Agentic terminal coding</td>
<td>Terminus-2 harness</td>
<td style="background: #e6f3ff">47.6%</td>
<td><strong>54.2%</strong></td>
<td>16.9%</td>
<td>32.6%</td>
<td>42.8%</td>
<td>—</td>
<td>—</td>
</tr>
<tr class="even">
<td>SWE-bench Verified</td>
<td>Agentic coding</td>
<td>Single attempt</td>
<td style="background: #e6f3ff">78.0%</td>
<td>76.2%</td>
<td>60.4%</td>
<td>59.6%</td>
<td>77.2%</td>
<td><strong>80.0%</strong></td>
<td>50.6%</td>
</tr>
<tr class="odd">
<td>τ2-bench</td>
<td>Agentic tool use</td>
<td></td>
<td style="background: #e6f3ff">90.2%</td>
<td><strong>90.7%</strong></td>
<td>79.5%</td>
<td>77.8%</td>
<td>87.2%</td>
<td>—</td>
<td>—</td>
</tr>
<tr class="even">
<td>Toolathlon</td>
<td>Long horizon real-world software tasks</td>
<td></td>
<td style="background: #e6f3ff"><strong>49.4%</strong></td>
<td>36.4%</td>
<td>3.7%</td>
<td>10.5%</td>
<td>38.9%</td>
<td>46.3%</td>
<td>—</td>
</tr>
<tr class="odd">
<td>MCP Atlas</td>
<td>Multi-step workflows using MCP</td>
<td></td>
<td style="background: #e6f3ff">57.4%</td>
<td>54.1%</td>
<td>3.4%</td>
<td>8.8%</td>
<td>43.8%</td>
<td><strong>60.6%</strong></td>
<td>—</td>
</tr>
<tr class="even">
<td>Vending-Bench 2</td>
<td>Agentic long term coherence</td>
<td>Net worth (mean),<br />
higher is better</td>
<td style="background: #e6f3ff">$3,635</td>
<td><strong>$5,478</strong></td>
<td>$549</td>
<td>$574</td>
<td>$3,839</td>
<td>$3,952</td>
<td>$1,107</td>
</tr>
<tr class="odd">
<td>FACTS Benchmark Suite</td>
<td>Factuality benchmark across<br />
grounding, parametric, search,<br />
and MM</td>
<td></td>
<td style="background: #e6f3ff">61.9%</td>
<td><strong>70.5%</strong></td>
<td>50.4%</td>
<td>63.4%</td>
<td>48.9%</td>
<td>61.4%</td>
<td>42.1%</td>
</tr>
<tr class="even">
<td>SimpleQA Verified</td>
<td>Parametric knowledge</td>
<td></td>
<td style="background: #e6f3ff">68.7%</td>
<td><strong>72.1%</strong></td>
<td>28.1%</td>
<td>54.5%</td>
<td>29.3%</td>
<td>38.0%</td>
<td>19.5%</td>
</tr>
<tr class="odd">
<td>MMMLU</td>
<td>Multilingual Q&amp;A</td>
<td></td>
<td style="background: #e6f3ff"><strong>91.8%</strong></td>
<td><strong>91.8%</strong></td>
<td>86.6%</td>
<td>89.5%</td>
<td>89.1%</td>
<td>89.6%</td>
<td>86.8%</td>
</tr>
<tr class="even">
<td>Global PIQA</td>
<td>Commonsense reasoning across<br />
100 Languages and Cultures</td>
<td></td>
<td style="background: #e6f3ff">92.8%</td>
<td><strong>93.4%</strong></td>
<td>90.2%</td>
<td>91.5%</td>
<td>90.1%</td>
<td>91.2%</td>
<td>85.6%</td>
</tr>
<tr class="odd">
<td>MRCR v2 (8-needle)</td>
<td>Long context performance</td>
<td>128K (average)<br />
1M (pointwise)</td>
<td style="background: #e6f3ff">67.2%<br />
22.1%</td>
<td>77.0%<br />
<strong>26.3%</strong></td>
<td>54.3%<br />
21.0%</td>
<td>58.0%<br />
16.4%</td>
<td>47.1%<br />
not supported</td>
<td><strong>81.9%</strong><br />
not supported</td>
<td>54.6%<br />
6.1%</td>
</tr>
</tbody>
</table>