-
Notifications
You must be signed in to change notification settings - Fork 29
Expand file tree
/
Copy pathindex.html
More file actions
1 lines (1 loc) · 31.3 KB
/
index.html
File metadata and controls
1 lines (1 loc) · 31.3 KB
1
<!doctype html><html lang="zh"><head><meta charset="UTF-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Step3-VL-10B: Compact Yet Frontier Multimodal Intelligence</title><script src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js" async></script><script type="module" crossorigin src="https://research-cdn.stepfun.ai/step3-vl-10b/js/main-B3DUhOn6.js"></script><link rel="stylesheet" href="https://research-cdn.stepfun.ai/step3-vl-10b/css/main-DSDtJ68C.css"></head><body class="theme-a"><header class="header"><div class="header-inner"><div class="logo-group"><img src="https://research-cdn.stepfun.ai/step3-vl-10b/images/logo-CnfAFL2t.png" alt="Step3-VL-10B" class="logo"></div><nav class="nav"><a href="#benchmark" data-i18n="nav.benchmark">Benchmark</a> <a href="#showcase" data-i18n="nav.showcase">Showcase</a> <a href="#method" data-i18n="nav.method">Method</a></nav><div class="header-actions"><button class="btn-ghost" id="theme-toggle" onclick="toggleTheme()"><span id="theme-label">🌙</span></button> <button class="btn-ghost" id="lang-toggle" onclick="toggleLang()"><span id="lang-label">EN</span></button></div></div></header><main class="main"><section id="hero" class="hero"><div class="container"><div class="hero-header"><h1 class="hero-title-large"><span class="title-static">Step3-VL-10B: </span><span class="title-typed" id="typed-text"></span><span class="typed-cursor">|</span></h1><p class="hero-meta">10B Parameters · 2026-01 · StepFun</p><div class="hero-links"><a href="https://github.com/stepfun-ai/Step3-VL-10B" class="link-item"><svg width="20" height="20" viewBox="0 0 24 24" fill="currentColor"><path d="M12 0C5.37 0 0 5.37 0 12c0 5.31 3.435 9.795 8.205 11.385.6.105.825-.255.825-.57 0-.285-.015-1.23-.015-2.235-3.015.555-3.795-.735-4.035-1.41-.135-.345-.72-1.41-1.23-1.695-.42-.225-1.02-.78-.015-.795.945-.015 1.62.87 1.845 1.23 1.08 1.815 2.805 1.305 3.495.99.105-.78.42-1.305.765-1.605-2.67-.3-5.46-1.335-5.46-5.925 0-1.305.465-2.385 1.23-3.225-.12-.3-.54-1.53.12-3.18 0 0 1.005-.315 3.3 1.23.96-.27 1.98-.405 3-.405s2.04.135 3 .405c2.295-1.56 3.3-1.23 3.3-1.23.66 1.65.24 2.88.12 3.18.765.84 1.23 1.905 1.23 3.225 0 4.605-2.805 5.625-5.475 5.925.435.375.81 1.095.81 2.22 0 1.605-.015 2.895-.015 3.3 0 .315.225.69.825.57A12.02 12.02 0 0024 12c0-6.63-5.37-12-12-12z"/></svg> GitHub </a><a href="https://huggingface.co/stepfun-ai/Step3-VL-10B" class="link-item"><svg width="20" height="20" viewBox="0 0 24 24" fill="currentColor"><path d="M12 2L2 7l10 5 10-5-10-5zM2 17l10 5 10-5M2 12l10 5 10-5"/></svg> HuggingFace </a><a href="https://modelscope.cn/models/stepfun-ai/Step3-VL-10B" class="link-item"><svg width="20" height="20" viewBox="0 0 24 24" fill="currentColor"><path d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm-1 17.93c-3.95-.49-7-3.85-7-7.93 0-.62.08-1.21.21-1.79L9 15v1c0 1.1.9 2 2 2v1.93zm6.9-2.54c-.26-.81-1-1.39-1.9-1.39h-1v-3c0-.55-.45-1-1-1H8v-2h2c.55 0 1-.45 1-1V7h2c1.1 0 2-.9 2-2v-.41c2.93 1.19 5 4.06 5 7.41 0 2.08-.8 3.97-2.1 5.39z"/></svg> ModelScope </a><a href="" class="link-item"><svg width="20" height="20" viewBox="0 0 24 24" fill="currentColor"><path d="M19 3H5c-1.1 0-2 .9-2 2v14c0 1.1.9 2 2 2h14c1.1 0 2-.9 2-2V5c0-1.1-.9-2-2-2zm-5 14H7v-2h7v2zm3-4H7v-2h10v2zm0-4H7V7h10v2z"/></svg> Paper</a></div></div><div class="teaser-chart-container"><div class="teaser-chart-header"><h2 class="teaser-chart-title">Frontier Performance, Minimal Cost</h2><p class="teaser-chart-subtitle">Average Benchmark Score vs Model Size</p><p class="teaser-chart-note">Average of: MMMU, MathVision, MathVista, MMBench (EN), MMBench (CN)</p></div><div class="teaser-chart-wrapper"><div class="teaser-chart-area" id="teaserChartArea"><div class="teaser-y-axis"><div class="teaser-axis-label">Avg Score</div><div class="teaser-axis-ticks" id="teaserYTicks"></div></div><div class="teaser-plot-area" id="teaserPlotArea"><div class="teaser-grid-lines" id="teaserGridLines"></div><div class="teaser-closed-ref-lines" id="teaserClosedRefLines"></div><div class="teaser-data-points" id="teaserDataPoints"></div></div><div class="teaser-x-axis"><div class="teaser-axis-ticks" id="teaserXTicks"></div><div class="teaser-axis-label">Parameters (B)</div></div></div></div><div class="teaser-chart-legend"><div class="teaser-legend-item teaser-legend-highlight"><div class="teaser-legend-shape"><div class="teaser-legend-square highlight"></div></div><span>Step3-VL-10B (SeRe)</span></div><div class="teaser-legend-item teaser-legend-pacore"><div class="teaser-legend-shape"><div class="teaser-legend-square pacore"></div></div><span>Step3-VL-10B (PaCoRe)</span></div><div class="teaser-legend-item"><div class="teaser-legend-shape"><div class="teaser-legend-square"></div></div><span>7-10B Models</span></div><div class="teaser-legend-item"><div class="teaser-legend-shape"><div class="teaser-legend-circle"></div></div><span>Flagship Models</span></div></div></div><div class="abstract-block"><p class="paragraph" data-i18n="abstract.p1"><strong>Step3-VL-10B</strong> 是一款轻量级开源基础模型，旨在重新定义<em>紧凑高效</em>与<em>前沿多模态智能</em>之间的权衡。尽管仅有 <strong>10B 参数</strong>，STEP3-VL-10B 在<em>视觉感知</em>、<em>复杂推理</em>和<em>人类对齐</em>方面表现卓越。</p><p class="paragraph" data-i18n="abstract.p2">该模型在 10B 规模以下的模型中始终表现最优，并能媲美甚至超越规模大 <em>10×–20×</em> 的开源模型（如 GLM-4.6V 106B-A12B、Qwen3-VL-Thinking 235B-A22B）以及顶级闭源旗舰模型（如 Gemini 2.5 Pro、Seed-1.5-VL）。</p><p class="paragraph" data-i18n="abstract.p3">Step3-VL-10B 的成功源于两大核心设计：<em>高质量多模态语料库的统一预训练</em>（1.2T tokens）与<em>规模化多模态强化学习</em>（超过 1,400 次 RL 迭代），并引入 <em>Parallel Coordinated Reasoning (PaCoRe)</em> 实现并行视觉探索的证据聚合。</p></div></div></section><section id="benchmark" class="section"><div class="container-full"><h2 class="section-title" data-i18n="section.benchmark">Benchmark</h2><div class="bar-charts-container"><div class="bar-charts-header"><h3 class="bar-charts-title">Benchmark Performance by Category</h3><p class="bar-charts-subtitle">Performance across MMMU, MathVista, MathVision, MMBench, AIME2025, and MultiChallenge benchmarks</p></div><div class="bar-charts-stack"><div class="bar-chart-panel" id="mmmuChart"><div class="bar-panel-header"><h4 class="bar-panel-title">MMMU</h4><span class="bar-panel-badge">Multimodal</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="mmmuYAxis"></div><div class="bar-plot" id="mmmuPlot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div><div class="bar-chart-panel" id="mathVistaChart"><div class="bar-panel-header"><h4 class="bar-panel-title">MathVista</h4><span class="bar-panel-badge">Multimodal</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="mathVistaYAxis"></div><div class="bar-plot" id="mathVistaPlot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div><div class="bar-chart-panel" id="mathVisionChart"><div class="bar-panel-header"><h4 class="bar-panel-title">MathVision</h4><span class="bar-panel-badge">Multimodal</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="mathVisionYAxis"></div><div class="bar-plot" id="mathVisionPlot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div><div class="bar-chart-panel" id="mmbenchChart"><div class="bar-panel-header"><h4 class="bar-panel-title">MMBench</h4><span class="bar-panel-badge">Multimodal</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="mmbenchYAxis"></div><div class="bar-plot" id="mmbenchPlot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div><div class="bar-chart-panel" id="aime2025Chart"><div class="bar-panel-header"><h4 class="bar-panel-title">AIME2025</h4><span class="bar-panel-badge">Text</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="aime2025YAxis"></div><div class="bar-plot" id="aime2025Plot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div><div class="bar-chart-panel" id="multiChallengeChart"><div class="bar-panel-header"><h4 class="bar-panel-title">MultiChallenge</h4><span class="bar-panel-badge">Text</span></div><div class="bar-chart-area"><div class="bar-y-axis" id="multiChallengeYAxis"></div><div class="bar-plot" id="multiChallengePlot"><div class="bar-section-label bar-open-label">7-10B Models</div><div class="bar-section-label bar-closed-label">Flagship Models</div></div></div></div></div><div class="bar-charts-legend"><div class="bar-legend-section"><span class="bar-legend-label">7-10B Models:</span><div class="bar-legend-item bar-highlight"><span class="bar-legend-bar"></span> <span>Step3-VL-10B (SeRe)</span></div><div class="bar-legend-item bar-pacore"><span class="bar-legend-bar bar-pacore"></span> <span>Step3-VL-10B (PaCoRe)</span></div><div class="bar-legend-item"><span class="bar-legend-bar bar-open"></span> <span>Others</span></div></div><div class="bar-legend-divider"></div><div class="bar-legend-section"><span class="bar-legend-label">Flagship Models:</span><div class="bar-legend-item bar-closed"><span class="bar-legend-bar bar-closed"></span> <span>Gemini 2.5 Pro / Seed-1.5-VL / GLM-4.6V (106B-A12B) / Qwen3-VL (235B-A22B)</span></div></div></div></div><p class="section-intro" data-i18n="bmk.intro">评测采用"<em>STEM 推理、识别、OCR & 文档、GUI Grounding、空间理解、代码</em>"等核心维度，以横向对比方式呈现多个同行模型的分数差异。对比表格强调<em>统计口径一致性</em>：同一数据集版本、统一评测脚本、固定温度与采样参数。</p><div class="bmk-table-wrap"><table class="bmk-table"><thead><tr><th class="bmk-th-benchmark" rowspan="2">Benchmark</th><th class="bmk-th-air" colspan="2">STEP3-VL-10B</th><th rowspan="2">GLM-4.6V</th><th rowspan="2">Qwen3-VL</th><th rowspan="2">Gemini-2.5 Pro</th><th rowspan="2">Seed-1.5-VL</th></tr><tr><th class="bmk-th-air bmk-th-sere">SeRe</th><th class="bmk-th-air bmk-th-pacore">PaCoRe</th></tr><tr class="bmk-row-params"><td></td><td class="bmk-col-air bmk-col-sere"><em>10B</em></td><td class="bmk-col-air bmk-col-pacore"><em>10B</em></td><td><em>106B-A12B</em></td><td><em>235B-A22B</em></td><td><em>—</em></td><td><em>—</em></td></tr></thead><tbody><tr class="bmk-row-divider"><td colspan="7">STEM / Multimodal Reasoning</td></tr><tr><td>MMMU</td><td class="bmk-col-sere">78.11</td><td class="bmk-col-pacore">80.11</td><td>75.20</td><td>78.70</td><td><strong>83.89</strong></td><td>79.11</td></tr><tr><td>MMMU-Pro</td><td class="bmk-col-sere">64.08</td><td class="bmk-col-pacore">67.18</td><td>65.84</td><td>72.37</td><td><strong>76.96</strong></td><td>70.60</td></tr><tr><td>MathVision</td><td class="bmk-col-sere">70.81</td><td class="bmk-col-pacore"><strong>75.95</strong></td><td>63.50*</td><td>72.10</td><td>73.30*</td><td>68.70*</td></tr><tr><td>MathVista</td><td class="bmk-col-sere">83.97</td><td class="bmk-col-pacore">85.50</td><td>83.51</td><td>85.10</td><td>83.88</td><td><strong>85.60</strong></td></tr><tr><td>LogicVista</td><td class="bmk-col-sere">66.89</td><td class="bmk-col-pacore">71.36</td><td>64.88</td><td><strong>73.15</strong></td><td>69.80</td><td>72.93</td></tr><tr><td>DynaMath</td><td class="bmk-col-sere">56.39</td><td class="bmk-col-pacore"><strong>61.48</strong></td><td>56.29</td><td>60.30</td><td>52.30</td><td>58.88</td></tr><tr><td>ZeroBench (main)</td><td class="bmk-col-sere">1.00</td><td class="bmk-col-pacore"><strong>5.00</strong></td><td>1.00</td><td>3.00</td><td>4.00</td><td>1.00</td></tr><tr><td>ZeroBench (sub)</td><td class="bmk-col-sere">27.54</td><td class="bmk-col-pacore">29.94</td><td>29.04</td><td>28.40</td><td><strong>33.53</strong></td><td>31.74</td></tr><tr><td>MathVerse (vision)</td><td class="bmk-col-sere">75.73</td><td class="bmk-col-pacore"><strong>78.30</strong></td><td>72.84</td><td>76.65</td><td>78.30</td><td>77.79</td></tr><tr><td>We-Math</td><td class="bmk-col-sere">73.03</td><td class="bmk-col-pacore">73.90</td><td>71.14</td><td>74.70</td><td><strong>80.10</strong></td><td>79.05</td></tr><tr><td>VisuLogic</td><td class="bmk-col-sere">29.68</td><td class="bmk-col-pacore">32.70</td><td>28.30</td><td>31.80</td><td>31.40</td><td><strong>34.30</strong></td></tr><tr><td>PhyX</td><td class="bmk-col-sere">59.45</td><td class="bmk-col-pacore">66.01</td><td>59.70</td><td>66.30</td><td><strong>67.56</strong></td><td>62.53</td></tr><tr class="bmk-row-divider"><td colspan="7">Recognition / General VQA</td></tr><tr><td>MMBench (EN)</td><td class="bmk-col-sere">92.05</td><td class="bmk-col-pacore">92.38</td><td>92.75</td><td>92.70</td><td><strong>93.19</strong></td><td>92.11</td></tr><tr><td>MMBench (CN)</td><td class="bmk-col-sere">91.55</td><td class="bmk-col-pacore">91.96</td><td>91.88</td><td>91.80</td><td><strong>93.13</strong></td><td>91.76</td></tr><tr><td>SimpleVQA</td><td class="bmk-col-sere">53.08</td><td class="bmk-col-pacore">54.64</td><td>57.95</td><td>59.30</td><td><strong>66.85</strong></td><td>64.72</td></tr><tr><td>MMStar</td><td class="bmk-col-sere">77.48</td><td class="bmk-col-pacore">77.64</td><td>75.30</td><td>76.80</td><td><strong>79.18</strong></td><td>77.91</td></tr><tr><td>HallusionBench</td><td class="bmk-col-sere">64.91</td><td class="bmk-col-pacore">64.54</td><td>60.63</td><td>65.58</td><td><strong>65.63</strong></td><td>64.13</td></tr><tr><td>MMVP</td><td class="bmk-col-sere">68.16</td><td class="bmk-col-pacore">68.00</td><td>71.33</td><td>71.30</td><td>70.67</td><td><strong>74.00</strong></td></tr><tr><td>ReMI</td><td class="bmk-col-sere">67.29</td><td class="bmk-col-pacore">69.12</td><td>64.42</td><td><strong>74.70</strong></td><td>71.69</td><td>72.19</td></tr><tr><td>M3GIA</td><td class="bmk-col-sere">78.33</td><td class="bmk-col-pacore">73.50</td><td>78.72</td><td>81.00</td><td>83.11</td><td><strong>83.22</strong></td></tr><tr><td>DoYouSeeMe</td><td class="bmk-col-sere">67.48</td><td class="bmk-col-pacore">68.54</td><td>67.50</td><td><strong>72.89</strong></td><td>71.19</td><td>71.94</td></tr><tr class="bmk-row-divider"><td colspan="7">Counting</td></tr><tr><td>CountBench</td><td class="bmk-col-sere">88.75</td><td class="bmk-col-pacore">88.80</td><td>92.06</td><td><strong>92.46</strong></td><td>87.78</td><td>91.85</td></tr><tr><td>CountQA</td><td class="bmk-col-sere">33.69</td><td class="bmk-col-pacore">38.29</td><td>36.32</td><td>45.62</td><td>38.02</td><td><strong>48.89</strong></td></tr><tr><td>PixMo-Count</td><td class="bmk-col-sere">70.85</td><td class="bmk-col-pacore">71.61</td><td>76.47</td><td>79.80</td><td>75.54</td><td><strong>83.38</strong></td></tr><tr class="bmk-row-divider"><td colspan="7">OCR</td></tr><tr><td>OCRBench</td><td class="bmk-col-sere">86.75</td><td class="bmk-col-pacore"><strong>89.00</strong></td><td>86.20</td><td>87.30</td><td>85.90</td><td>85.20</td></tr><tr><td>OmniOCR</td><td class="bmk-col-sere">76.98</td><td class="bmk-col-pacore">78.14</td><td>84.53</td><td>87.20</td><td>66.05</td><td><strong>87.80</strong></td></tr><tr><td>CC-OCR (Multi-Lang-OCR)</td><td class="bmk-col-sere">76.59</td><td class="bmk-col-pacore">77.51</td><td>74.08</td><td>80.80</td><td><strong>81.10</strong></td><td>78.82</td></tr><tr class="bmk-row-divider"><td colspan="7">2D / 3D Spatial Understanding</td></tr><tr><td>BLINK</td><td class="bmk-col-sere">66.79</td><td class="bmk-col-pacore">67.39</td><td>68.17</td><td>67.12</td><td><strong>72.01</strong></td><td>71.54</td></tr><tr><td>CVBench</td><td class="bmk-col-sere">83.49</td><td class="bmk-col-pacore">85.92</td><td>83.72</td><td><strong>87.86</strong></td><td>84.36</td><td>86.27</td></tr><tr><td>MMSI-Bench</td><td class="bmk-col-sere">32.18</td><td class="bmk-col-pacore">36.40</td><td>30.80</td><td>32.50</td><td><strong>40.40</strong></td><td>30.60</td></tr><tr><td>ERQA</td><td class="bmk-col-sere">48.87</td><td class="bmk-col-pacore">51.75</td><td>47.75</td><td>53.50</td><td><strong>62.25</strong></td><td>48.50</td></tr><tr><td>OmniSpatial</td><td class="bmk-col-sere">51.58</td><td class="bmk-col-pacore">52.58</td><td>50.49</td><td>53.10</td><td><strong>55.64</strong></td><td>51.99</td></tr><tr><td>All-Angles-Bench</td><td class="bmk-col-sere">57.21</td><td class="bmk-col-pacore">64.71</td><td>62.94</td><td>60.59</td><td><strong>65.88</strong></td><td>57.65</td></tr><tr><td>MindCube-tiny</td><td class="bmk-col-sere">62.81</td><td class="bmk-col-pacore">68.58</td><td>52.83</td><td>47.58</td><td>58.92</td><td><strong>39.83</strong></td></tr><tr><td>RealWorldQA</td><td class="bmk-col-sere">74.44</td><td class="bmk-col-pacore">75.56</td><td>77.78</td><td>78.80</td><td>77.78</td><td><strong>79.61</strong></td></tr><tr><td>SpatialViz-Bench</td><td class="bmk-col-sere">45.51</td><td class="bmk-col-pacore"><strong>52.03</strong></td><td>37.46</td><td>46.36</td><td>45.34</td><td>35.25</td></tr><tr><td>STARE</td><td class="bmk-col-sere">61.75</td><td class="bmk-col-pacore">64.57</td><td>60.38</td><td><strong>70.89</strong></td><td>62.36</td><td>62.99</td></tr><tr><td>CoreCognition</td><td class="bmk-col-sere">66.69</td><td class="bmk-col-pacore">71.54</td><td>69.50</td><td>72.66</td><td><strong>78.78</strong></td><td>72.38</td></tr><tr><td>V*</td><td class="bmk-col-sere">82.85</td><td class="bmk-col-pacore">84.29</td><td>85.86</td><td>89.53</td><td>80.63</td><td><strong>90.58</strong></td></tr><tr><td>ViewSpatial</td><td class="bmk-col-sere">46.14</td><td class="bmk-col-pacore">48.41</td><td>43.87</td><td><strong>48.58</strong></td><td>44.15</td><td>44.14</td></tr><tr class="bmk-row-divider"><td colspan="7">Exam (Text-Centric)</td></tr><tr><td>MMLU-Pro</td><td class="bmk-col-sere">76.02</td><td class="bmk-col-pacore">77.09</td><td>79.96</td><td>83.75</td><td><strong>86.45</strong></td><td>83.39</td></tr><tr><td>GPQA-Diamond</td><td class="bmk-col-sere">70.83</td><td class="bmk-col-pacore">73.99</td><td>69.19</td><td>77.68</td><td><strong>84.06</strong></td><td>71.91</td></tr><tr><td>SuperGPQA</td><td class="bmk-col-sere">50.38</td><td class="bmk-col-pacore">53.15</td><td>53.28</td><td>64.20</td><td><strong>65.00</strong></td><td>60.50</td></tr><tr><td>LiveBench (2024-11-25)</td><td class="bmk-col-sere">69.71</td><td class="bmk-col-pacore">71.69</td><td>62.75</td><td><strong>80.14</strong></td><td>76.34</td><td>65.62</td></tr><tr class="bmk-row-divider"><td colspan="7">Mathematics (Text-Centric)</td></tr><tr><td>AIME 2024</td><td class="bmk-col-sere">90.94</td><td class="bmk-col-pacore"><strong>93.33</strong></td><td>80.63</td><td>91.93</td><td>79.53</td><td>79.48</td></tr><tr><td>AIME 2025</td><td class="bmk-col-sere">87.66</td><td class="bmk-col-pacore"><strong>94.43</strong></td><td>71.88</td><td>83.59</td><td>83.96</td><td>64.06</td></tr><tr><td>HMMT 2025</td><td class="bmk-col-sere">78.18</td><td class="bmk-col-pacore"><strong>92.14</strong></td><td>57.29</td><td>67.71</td><td>65.68</td><td>51.30</td></tr><tr><td>CNMO 2024</td><td class="bmk-col-sere">78.20</td><td class="bmk-col-pacore">81.17</td><td>72.11</td><td><strong>88.36</strong></td><td>74.53</td><td>83.67</td></tr><tr><td>Beyond AIME</td><td class="bmk-col-sere">63.23</td><td class="bmk-col-pacore"><strong>74.00</strong></td><td>39.83</td><td>57.42</td><td>54.45</td><td>42.83</td></tr><tr><td>IMO-AnswerBench</td><td class="bmk-col-sere">62.12</td><td class="bmk-col-pacore"><strong>76.66</strong></td><td>51.25</td><td>69.25</td><td>72.00</td><td>44.75</td></tr><tr class="bmk-row-divider"><td colspan="7">Code</td></tr><tr><td>LiveCodeBench (2408-2505)</td><td class="bmk-col-sere">75.77</td><td class="bmk-col-pacore"><strong>76.43</strong></td><td>48.71</td><td>69.45</td><td>72.01</td><td>57.10</td></tr></tbody></table></div><p class="bmk-note" data-i18n="bmk.note.detail">注：<strong>SeRe</strong> (Sequential Reasoning) 使用最大 64K tokens；<strong>PaCoRe</strong> (Parallel Coordinated Reasoning) 聚合 16 个并行 rollouts，最大 128K tokens。测试采用: temperature=1, top_p=1, top_k=0。</p></div></section><div class="divider"></div><section id="showcase" class="section"><div class="container-wide"><h2 class="section-title" data-i18n="section.showcase">Showcase</h2><p class="section-intro" data-i18n="showcase.intro">Showcase 通过真实案例展示 Step3-VL-10B 的多模态推理能力：Case 1 聚焦莫尔斯电码表格解析，其他案例覆盖GUI感知与视觉识别和推理。</p><div class="carousel-nav-tabs"><button class="carousel-nav-btn active" data-index="0" data-i18n="showcase.cat1">2.1 莫尔斯推理</button> <button class="carousel-nav-btn" data-index="1" data-i18n="showcase.cat2">2.2 GUI 感知</button> <button class="carousel-nav-btn" data-index="2" data-i18n="showcase.cat3">2.3 视觉识别</button> <button class="carousel-nav-btn" data-index="3" data-i18n="showcase.cat4">2.4 图推理</button></div><div class="carousel-wrapper"><div class="carousel-track"><div class="carousel-item" data-index="0" data-pos="0"><div class="carousel-item-header"><div class="carousel-item-title" data-i18n="showcase.cat1">2.1 莫尔斯推理 / Morse Reasoning</div><div class="case-pill">CASE 1</div></div><div class="carousel-demo-user"><div class="carousel-demo-label">User</div><div class="case-highlight-media"><img src="https://research-cdn.stepfun.ai/step3-vl-10b/images/case1-wij6vhjT.png" alt="Morse code reference table for STEPFUN" loading="lazy"></div><div class="carousel-demo-text" data-i18n="demo.image.a.user">根据该表生成"STEPFUN"的编码</div></div><div class="carousel-demo-agent"><p class="case-highlight-label">Thinking Process</p><p class="case-highlight-text">用户现在需要把每个字母转换成莫尔斯电码，先看每个字母在表中的编码。首先分解"STEPFUN"每个字母：S、T、E、P、F、U、N。</p><ul class="case-highlight-list"><li><strong>S</strong>：看右边第三列第一个，S是<span class="morse-inline">···</span>（三个点）。（一个短横）</li><li><strong>T</strong>：第三列第二个，T是<span class="morse-inline">-</span>。</li><li><strong>E</strong>：左边第一列第五个，E是<span class="morse-inline">·</span>（一个点）。</li><li><strong>P</strong>：中间列（J到R）中P对应的符号是<span class="morse-inline">·--·</span>点、短横、短横、点）。</li><li><strong>F</strong>：最左侧列第六行，F的符号是<span class="morse-inline">··-·</span>（两个点、一个短横、一个点）。</li><li><strong>U</strong>：最右侧列第三行，U的符号是<span class="morse-inline">··-</span>（两个点、一个短横）。</li><li><strong>N</strong>：中间列（J到R）第五行，N的符号是<span class="morse-inline">-·</span>一个短横、一个点）。</li></ul><p class="case-highlight-text">再将每个字母的编码按顺序拼接（字母间用空格分隔更清晰），"STEPFUN"的莫尔斯电码为：<span class="morse-token">···</span> <span class="morse-token">-</span> <span class="morse-token">·</span> <span class="morse-token">·--·</span> <span class="morse-token">··-·</span> <span class="morse-token">··-</span> <span class="morse-token">-·</span></p></div><div class="case-highlight-section"><p class="case-highlight-label">Final Answer</p><div class="case-highlight-answer"><span class="morse-token">···</span> <span class="morse-token">-</span> <span class="morse-token">·</span> <span class="morse-token">·--·</span> <span class="morse-token">··-·</span> <span class="morse-token">··-</span> <span class="morse-token">-·</span></div></div></div><div class="carousel-item" data-index="1" data-pos="1"><div class="carousel-item-header"><div class="carousel-item-title" data-i18n="showcase.cat2">2.2 GUI 感知 / GUI Perception</div><div class="case-pill">CASE 2</div></div><div class="carousel-demo-user"><div class="carousel-demo-label">User</div><div class="case-highlight-media"><img src="https://research-cdn.stepfun.ai/step3-vl-10b/images/case2-BPTN5a1y.png" alt="LaTeX editor tabs for dissertation chapters" loading="lazy"></div><div class="carousel-demo-text">The screenshot shows a snapshot of a LaTeX text editor a student uses for writing their PhD dissertation. What is the minimum number of chapter .tex files the student has opened?</div></div><div class="carousel-demo-agent"><p class="case-highlight-label">Thinking Process</p><p class="case-highlight-text">So, let's look at the tabs at the top of the editor.</p><ul class="case-highlight-list"><li>chapte... (leftmost, maybe chapter1.tex as the active file in the center)</li><li>chapte... (next)</li><li>thesis... (this is thesis.tex, not a chapter)</li><li>chapte... (next)</li><li>chapte... (next)</li><li>chapte... (last one before the dark theme part)</li></ul><p class="case-highlight-text">To determine the minimum number of chapter .tex files opened, we analyze the tabs at the top of the editor.</p><ol class="case-highlight-list"><li>Identify tabs: The image shows multiple tabs labeled chapte... (truncated for "chapter...") and one tab labeled thesis....</li><li>Exclude non-chapter tabs: The thesis... tab is not a chapter file, so we count only the chapte... tabs.</li><li>Count chapter tabs: There are 5 visible chapte... tabs (excluding the thesis... tab).</li></ol></div><div class="case-highlight-section"><p class="case-highlight-label">Final Answer</p><div class="case-highlight-answer">5</div></div></div><div class="carousel-item" data-index="2" data-pos="2"><div class="carousel-item-header"><div class="carousel-item-title" data-i18n="showcase.cat3">2.3 视觉识别 / Visual Recognition</div><div class="case-pill">CASE 3</div></div><div class="carousel-demo-user"><div class="carousel-demo-label">User</div><div class="case-highlight-media"><img src="https://research-cdn.stepfun.ai/step3-vl-10b/images/case3-AwMsX3Af.png" alt="Out-of-focus monitor displaying HP logo" loading="lazy"></div><div class="carousel-demo-text">What is the brand name of the monitor in the image? What is its encoding with a Caesar cipher with a left shift of 5. Give your answer as a combined string.</div></div><div class="carousel-demo-agent"><p class="case-highlight-label">Thinking Process</p><p class="case-highlight-text">Got it, let's break this down. First, find the monitor's brand.</p><p class="case-highlight-text">...</p><p class="case-highlight-text">The logo on the monitor's base is HP. Then, we need to do a Caesar cipher with a left shift of 5.</p><p class="case-highlight-text">...</p><ul class="case-highlight-list"><li>First, H is the 8th letter (A=1, B=2... H=8). Left shift 5: 8-5=3 → C.</li><li>Then, P is the 16th letter. 16-5=11 → K.</li></ul><p class="case-highlight-text">So "HP" shifted left 5 is "CK".</p></div><div class="case-highlight-section"><p class="case-highlight-label">Final Answer</p><div class="case-highlight-answer">HPCK</div></div></div><div class="carousel-item" data-index="3" data-pos="-1"><div class="carousel-item-header"><div class="carousel-item-title" data-i18n="showcase.cat4">2.4 图推理 / Graph Reasoning</div><div class="case-pill">CASE 4</div></div><div class="carousel-demo-user"><div class="carousel-demo-label">User</div><div class="case-highlight-media"><img src="https://research-cdn.stepfun.ai/step3-vl-10b/images/case4-DUZzsUFW.png" alt="Weighted directed graph with labeled edges" loading="lazy"></div><div class="carousel-demo-text">How many edges have the shortest length?</div></div><div class="carousel-demo-agent"><p class="case-highlight-label">Thinking Process</p><p class="case-highlight-text">To determine the number of edges with the shortest length, we first identify the smallest edge weight in the diagram.</p><p class="case-highlight-text">...</p><p class="case-highlight-text">The smallest weight is 1. Next, we count all edges labeled with 1:</p><p class="case-highlight-text">...</p><ul class="case-highlight-list"><li><strong>A -> T</strong>: labeled 1</li><li><strong>T → X</strong>: labeled 1</li><li><strong>P → O</strong>: labeled 1</li><li><strong>W → X</strong>: labeled 1</li></ul><p class="case-highlight-text">...</p><p class="case-highlight-text">No other edges have a weight of 1. Thus, there are 4 edges with the shortest length.</p></div><div class="case-highlight-section"><p class="case-highlight-label">Final Answer</p><div class="case-highlight-answer">4</div></div></div></div></div><div class="carousel-controls"><button class="carousel-arrow" id="carousel-prev-btn">←</button><div class="carousel-dots"><div class="carousel-dot active" data-index="0"></div><div class="carousel-dot" data-index="1"></div><div class="carousel-dot" data-index="2"></div><div class="carousel-dot" data-index="3"></div></div><button class="carousel-arrow" id="carousel-next-btn">→</button></div></div></section><div class="divider"></div><section id="method" class="section"><div class="container"><h2 class="section-title" data-i18n="section.method">Method</h2><div class="method-content"><div class="formula-section"><h3 class="subsection-title" data-i18n="method.arch.title">Architecture</h3><ul class="case-highlight-list"><li data-i18n="method.arch.encoder">Visual Encoder: PE-lang (Language-Optimized Perception Encoder), 1.8B parameters.</li><li data-i18n="method.arch.decoder">Decoder: Qwen3-8B.</li><li data-i18n="method.arch.projector">Projector: Two consecutive stride-2 layers (resulting in 16× spatial downsampling).</li><li data-i18n="method.arch.resolution">Resolution: Multi-crop strategy consisting of a 728×728 global view and multiple 504×504 local crops.</li></ul></div><div class="formula-section"><h3 class="subsection-title" data-i18n="method.train.title">Training Pipeline</h3><p class="paragraph" data-i18n="method.train.pretrain"><strong>Pre-training:</strong> Single-stage, fully unfrozen strategy using AdamW optimizer (Total: 1.2T tokens, 370K iterations).</p><ul class="case-highlight-list"><li data-i18n="method.train.pretrain.p1">Phase 1: 900B tokens.</li><li data-i18n="method.train.pretrain.p2">Phase 2: 300B tokens.</li></ul><p class="paragraph" data-i18n="method.train.sft"><strong>Supervised Finetuning (SFT):</strong> Two-stage approach (Total: ~226B tokens).</p><ul class="case-highlight-list"><li data-i18n="method.train.sft.s1">Stage 1: 9:1 text-to-multimodal ratio (~190B tokens).</li><li data-i18n="method.train.sft.s2">Stage 2: 1:1 text-to-multimodal ratio (~36B tokens).</li></ul><p class="paragraph" data-i18n="method.train.rl"><strong>Reinforcement Learning:</strong> Total >1,400 iterations.</p><ul class="case-highlight-list"><li data-i18n="method.train.rl.rlvr">RLVR: 600 iterations (Tasks: mathematics, geometry, physics, perception, grounding).</li><li data-i18n="method.train.rl.rlhf">RLHF: 300 iterations (Task: open-ended generation).</li></ul></div></div></div></section><div class="divider"></div></main><footer class="site-footer"><div class="container"><div class="footer-bottom"><p class="copyright">© 2026 StepFun. All rights reserved.</p><div class="footer-links"><a href="#">Privacy</a> <a href="#">Terms</a> <a href="#">Contact</a></div></div></div></footer></body></html>