assets/jevvision-comparison-table.svg

<svg xmlns="http://www.w3.org/2000/svg" width="1800" height="1660" viewBox="0 0 1800 1660" role="img" aria-labelledby="title desc">
<title id="title">JEVision capability comparison</title>
<desc id="desc">A dark comparison chart of Jev, KEV-0.8B, and JEVision across output format, input modality, request context, image support, and JevBench text results. On the public panel JEVision is shown at 79.3 percent. On the general group of 33 questions it scores 21 of 33, or 63.6 percent.</desc>
<defs><style>
  text{font-family:Inter,Segoe UI,Arial,sans-serif;fill:#f1f2f1}
  .muted{fill:#a9adae}.quiet{fill:#818889}.accenttext{fill:#fa7954}
  .eyebrow{fill:#888f91;font-size:18px;font-weight:700;letter-spacing:3px}
  .rowlabel{fill:#a9adae;font-size:17px;font-weight:700;letter-spacing:2px}
  .head{font-size:28px;font-weight:700}.headsub{fill:#a9adae;font-size:17px}
  .cell{font-size:22px;font-weight:650}.cellsub{fill:#a9adae;font-size:18px}
  .cellstrong{font-size:18px;font-weight:650}.metric{font-size:30px;font-weight:720}
  .rule{stroke:#303536;stroke-width:1.5}
  .card{fill:#151819;stroke:#34393a;stroke-width:1.5}
  .note{fill:#1b1f20;stroke:#34393a;stroke-width:1.25}
</style></defs>
<rect width="1800" height="1660" fill="#0c0e0f"/>
<rect x="96" y="80" width="7" height="26" rx="3.5" fill="#fa7954"/>
<text x="121" y="101" class="eyebrow">JEVISION / AT A GLANCE</text>
<text x="96" y="204" font-size="70" font-weight="740" letter-spacing="-3">Jev-style decisions. A visual route.</text>
<text x="98" y="254" class="muted" font-size="27">A clear-eyed comparison of interface, context envelope and JevBench results.</text>
<line x1="96" y1="300" x2="1704" y2="300" class="rule"/>
<rect x="96" y="348" width="1608" height="1182" rx="18" class="card"/>
<rect x="96" y="348" width="1608" height="5" rx="2" fill="#fa7954"/>
<rect x="114" y="365" width="1572" height="70" rx="12" fill="#1c2021"/>
<rect x="1178" y="365" width="508" height="70" rx="12" fill="#2b211d" stroke="#a25742" stroke-width="1.5"/>
<text x="132" y="405" class="eyebrow">CAPABILITY</text>
<text x="324" y="397" class="head">Jev</text>
<text x="324" y="425" class="headsub">hosted TypeSafe API</text>
<text x="752" y="397" class="head">KEV-0.8B</text>
<text x="752" y="425" class="headsub">released text parent</text>
<text x="1200" y="397" class="head">JEVision</text>
<text x="1200" y="425" class="headsub">routed multimodal bundle</text>
<line x1="132" y1="450" x2="1668" y2="450" class="rule"/>
<text x="132" y="489" class="rowlabel">OUTPUT</text>
<text x="132" y="517" class="rowlabel">FORMAT</text>
<text x="324" y="493" class="cell">Typed decisions</text>
<text x="324" y="523" class="cellsub">Choice · Score · Noul</text>
<text x="752" y="493" class="cell">KEV-compatible</text>
<text x="752" y="523" class="cellsub">Choice · Score · Noul</text>
<text x="1200" y="493" class="cell">Jev-style typed answers</text>
<text x="1200" y="523" class="cellsub">Choice · Score · Noul</text>
<line x1="132" y1="553" x2="1668" y2="553" class="rule"/>
<text x="132" y="594" class="rowlabel">INPUT</text>
<text x="132" y="622" class="rowlabel">MODALITY</text>
<text x="324" y="597" class="cell">Text state</text>
<text x="324" y="627" class="cellsub">string, object or array</text>
<text x="752" y="597" class="cell">Text state</text>
<text x="752" y="627" class="cellsub">text-only serving route</text>
<text x="1200" y="597" class="cell">Text + images</text>
<text x="1200" y="627" class="cellsub">images take a separate visual route</text>
<line x1="132" y1="656" x2="1668" y2="656" class="rule"/>
<text x="132" y="708" class="rowlabel">REQUEST</text>
<text x="132" y="736" class="rowlabel">CONTEXT</text>
<text x="324" y="705" class="metric">64K</text>
<text x="324" y="742" class="cellstrong">per request</text>
<text x="324" y="774" class="cellsub">state + longest question ≤32K</text>
<text x="752" y="705" class="metric">8,192</text>
<text x="752" y="742" class="cellstrong">state tokens + 8,192 per question branch</text>
<text x="752" y="774" class="cellsub">serving limit; longer context untested</text>
<text x="1200" y="705" class="metric">80K</text>
<text x="1200" y="742" class="cellstrong">configured request cap</text>
<text x="1200" y="774" class="cellsub">not a 128K claim or quality guarantee</text>
<line x1="132" y1="827" x2="1668" y2="827" class="rule"/>
<text x="132" y="873" class="rowlabel">IMAGE</text>
<text x="132" y="901" class="rowlabel">SUPPORT</text>
<text x="324" y="870" class="cell">No raw-image field</text>
<text x="324" y="903" class="cellsub">in the documented API schema</text>
<text x="752" y="870" class="cell">Text only</text>
<text x="752" y="903" class="cellsub">in the KEV serving route</text>
<text x="1200" y="870" class="cell">Separate visual sidecar</text>
<text x="1200" y="903" class="cellsub">M3 text LoRA is not used on image calls;</text>
<text x="1200" y="930" class="cellsub">general image accuracy is not measured</text>
<line x1="132" y1="969" x2="1668" y2="969" class="rule"/>
<text x="132" y="1016" class="rowlabel">JEVBENCH</text>
<text x="132" y="1044" class="rowlabel">PUBLIC · n=231</text>
<text x="324" y="1017" class="metric">200 / 231</text>
<text x="324" y="1052" class="cellsub">86.6% · public reference</text>
<text x="752" y="1017" class="metric">139 / 231</text>
<text x="752" y="1052" class="cellsub">60.2% · KEV-0.8B parent</text>
<text x="1200" y="1017" class="metric accenttext">79.3%</text>
<text x="1200" y="1052" class="cellsub">M3 text adapter · public panel</text>
<line x1="132" y1="1099" x2="1668" y2="1099" class="rule"/>
<text x="132" y="1146" class="rowlabel">GENERAL GROUP</text>
<text x="132" y="1174" class="rowlabel">33 QUESTIONS</text>
<text x="324" y="1147" class="metric">29 / 33</text>
<text x="324" y="1182" class="cellsub">87.9% · reference</text>
<text x="752" y="1147" class="metric">17 / 33</text>
<text x="752" y="1182" class="cellsub">51.5% · KEV-0.8B parent</text>
<text x="1200" y="1147" class="metric">21 / 33</text>
<text x="1200" y="1182" class="cellsub">63.6% · below Jev on this group</text>
<line x1="132" y1="1229" x2="1668" y2="1229" class="rule"/>
<text x="132" y="1276" class="rowlabel">HOW TO</text>
<text x="132" y="1304" class="rowlabel">READ SCORES</text>
<text x="324" y="1277" class="cellstrong">Reference scores from the same</text>
<text x="324" y="1307" class="cellsub">benchmark report.</text>
<text x="752" y="1277" class="cellstrong">Parent checkpoint, not raw Qwen.</text>
<text x="752" y="1307" class="cellsub">Qwen3.5-0.8B-Base was not</text>
<text x="752" y="1334" class="cellsub">separately scored.</text>
<text x="1200" y="1277" class="cellstrong">Text-only benchmark scores.</text>
<text x="1200" y="1307" class="cellsub">Image calls use a separate route;</text>
<text x="1200" y="1334" class="cellsub">image quality is not assessed here.</text>
<line x1="132" y1="1396" x2="1668" y2="1396" class="rule"/>
<rect x="114" y="1414" width="1572" height="92" rx="14" class="note"/>
<circle cx="147" cy="1450" r="5.5" fill="#fa7954"/>
<text x="169" y="1455" class="cellstrong">80K is configured, not validated for accuracy.</text>
<text x="169" y="1482" class="cellsub">Earlier ~77K checks used the KEV-0.8B text route and visual sidecar separately.</text>
<line x1="96" y1="1550" x2="1704" y2="1550" class="rule"/>
<text x="96" y="1585" class="quiet" font-size="16">Sources: TypeSafe API schema · KEV README · JEVision release bundle · JevBench M3 v1.3 report</text>
<text x="96" y="1615" class="quiet" font-size="16">The general question group is separate from the public panel; all JevBench results shown are text-only.</text>
<text x="1704" y="1638" class="quiet" font-size="14" text-anchor="end">JEVISION · CAPABILITY COMPARISON · SEP 2026</text>
</svg>