---
title: Edge0-145B-NVFP4
canonical_url: "https://www.modelscope.cn/models/Edge0/Edge0-145B-NVFP4"
md_url: "https://www.modelscope.cn/models/Edge0/Edge0-145B-NVFP4.md"
repository: Edge0/Edge0-145B-NVFP4
last_updated: 2026-11-20
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
base_model:
  - Qwen/Qwen3.8-Flash-Next
base_model_relation: quantized
library_name:
  - gguf
  - pytorch
frameworks:
  - pytorch
downloads: 849
stars: 0
tags:
  - moe
  - edge-inference
  - latent-moe
  - 4-bit
  - agent
  - coding
  - cuda
---

# Edge0-145B-NVFP4

> Edge0-145B-NVFP4 - Edge0 在 ModelScope 开源的模型。.edge0-card :where(p,span,small,strong,b,summary,div,nav,li,td,th) {font-family:inherit;line-height:inherit;color:inherit;} .edge0-card{color-scheme:light; --ink:#1d1d1f; --muted:#6e6e73; --line:#e5e5e7;…

Edge0/Edge0-145B-NVFP4 是 ModelScope 魔搭社区上的text-generation模型，采用 other 许可，基于 Qwen/Qwen3.8-Flash-Next 构建。

- **Repository**: Edge0/Edge0-145B-NVFP4
- **License**: other
- **Tasks**: text-generation
- **Base model**: Qwen/Qwen3.8-Flash-Next
- **Tags**: moe, edge-inference, latent-moe, 4-bit, agent, coding, cuda
- **Downloads**: 849
- **Stars**: 0
- **Last updated**: 2026-11-20

Source: https://www.modelscope.cn/models/Edge0/Edge0-145B-NVFP4

---

<style>
.edge0-card :where(p,span,small,strong,b,summary,div,nav,li,td,th) {font-family:inherit;line-height:inherit;color:inherit;}
.edge0-card{color-scheme:light; --ink:#1d1d1f; --muted:#6e6e73; --line:#e5e5e7;
  --signal:#E63946; --ivory:#F1FAEE; --ice:#A8DADC; --steel:#45789D; --navy:#1D3557; --canvas:#ffffff; --wash:#f5f5f7;}
.edge0-card,.edge0-card *{box-sizing:border-box;}
.edge0-card{margin:0; background:var(--canvas); color:var(--ink);
  font-family:-apple-system,BlinkMacSystemFont,"SF Pro Text","Helvetica Neue",Inter,Arial,sans-serif;
  font-size:15px; line-height:1.8; overflow-wrap:break-word; -webkit-font-smoothing:antialiased;}
.edge0-card a{color:#515154; text-decoration:none; text-underline-offset:4px;}
.edge0-card a:hover{color:var(--ink); text-decoration:underline;}
.edge0-card a:focus-visible,.edge0-card summary:focus-visible,.edge0-card [tabindex]:focus-visible{outline:3px solid var(--steel); outline-offset:5px; border-radius:4px;}
.edge0-card > div[align="center"]{text-align:center;}
.edge0-card > div[align="center"] > p{margin:14px 0;}
.edge0-card h1{color:var(--ink); font-size:35px; font-weight:650; letter-spacing:-1px; line-height:1.28;
  margin:30px 0 19px;}
.edge0-card h2{color:var(--ink); font-size:25px; font-weight:600; letter-spacing:-.5px; line-height:1.35;
  margin:58px 0 22px; padding-bottom:13px; border-bottom:1px solid var(--line); scroll-margin-top:24px;}
.edge0-card h3{color:var(--ink); font-size:17px; font-weight:600; margin:28px 0 12px;}
.edge0-card p{margin:17px 0;}
.edge0-card b,.edge0-card strong{font-weight:600; color:var(--ink);}
.edge0-card img{max-width:100%; height:auto; border-radius:10px; vertical-align:middle;}
.edge0-card img.badge{width:auto; height:20px; max-width:100%; border-radius:0; margin:3px;}
.edge0-card .hero-image{display:block; width:100%; margin:0 auto 28px;}
.edge0-card .card-hero > p:has(> .hero-image){margin:0;}
.edge0-card .card-hero .hero-image{margin:0; border-radius:0;}
.edge0-card .card-hero > h1{color:var(--ink); font-size:24px; font-weight:500; letter-spacing:-.6px;
  line-height:1.3; margin:46px 0 0;}
.edge0-card .editorial-art{margin:0; line-height:0;}
.edge0-card .editorial-image{display:block; width:100%; max-width:720px; height:auto; margin:0 auto;}
.edge0-card .editorial-pair{position:relative; display:grid; grid-template-columns:repeat(2,minmax(0,1fr));
  align-items:start; gap:64px; max-width:720px; margin:0 auto; padding:64px 0 70px;}
.edge0-card .editorial-pair::after{content:""; position:absolute; left:50%; top:50%;
  width:1px; height:72px; background:#d2d2d7; transform:translate(-50%,-50%);}
.edge0-card .editorial-pillar{min-width:0; text-align:center;}
.edge0-card .editorial-pillar .pillar-title{max-width:240px; margin:0 auto 13px; color:var(--ink);
  font-size:25px; font-weight:500; letter-spacing:-.55px; line-height:1.25;}
.edge0-card .editorial-pillar .pillar-caption{max-width:205px; margin:0 auto; color:var(--muted);
  font-size:12px; line-height:1.7;}
.edge0-card .card-hero > p.badge-github{margin:0 0 8px; line-height:1;}
.edge0-card .card-hero > p.badge-models{display:flex; flex-wrap:nowrap; width:max-content;
  max-width:100%; margin:0 auto; padding:0 0 7px; overflow-x:auto; white-space:nowrap;
  line-height:1; scrollbar-width:thin; scrollbar-color:#d2d2d7 #fafafa;
  -webkit-overflow-scrolling:touch;}
.edge0-card .badge-models > a{flex:0 0 auto;}
.edge0-card .model-context{max-width:740px; margin:50px auto 48px;}
.edge0-card .model-context p{color:var(--muted); font-size:14px; line-height:1.9;}
.edge0-card .model-context strong{color:var(--ink); font-weight:500;}
.edge0-card .quality-figure-link{display:block;}
.edge0-card .quality-chart{display:block; width:100%; max-width:1400px; margin:30px auto 34px; border-radius:0;}
.edge0-card .repo-summary{margin:58px 0 38px;}
.edge0-card .repo-summary > p:first-child{margin:0 0 17px; font-size:14px;}
.edge0-card .repo-summary > p:first-child strong{white-space:nowrap;}
.edge0-card .repo-summary > p:last-child{margin:16px 0 0;}
.edge0-card .repo-note{display:block; position:relative; padding-left:19px; text-indent:-19px;
  font-size:12px; line-height:1.8; color:var(--muted);}
.edge0-card .repo-note strong{color:inherit; font-weight:500;}
.edge0-card > div[align="center"] > p:first-child{margin:0;}
.edge0-card > div[align="center"] .section-nav{margin:28px 0 36px; line-height:2.5;}
.edge0-card .section-nav a{display:inline-block; color:#6e6e73; font-size:12px; margin:0 6px;}
.edge0-card .section-nav a:hover{color:var(--ink);}
.edge0-card code{background:#f5f5f7; color:#515154; padding:2px 6px; border-radius:4px; font-size:87%;
  font-family:"SF Mono",Menlo,Consolas,monospace; overflow-wrap:anywhere;}
.edge0-card pre{background:#f5f5f7; color:var(--ink); padding:20px 22px; border:1px solid #e5e5e7;
  border-radius:8px; overflow-x:auto; max-width:100%; line-height:1.7; margin:22px 0;}
.edge0-card pre code{background:none; color:inherit; padding:0; font-size:13px; overflow-wrap:normal;}
.edge0-card .table-scroll{max-width:100%; overflow-x:auto; margin:25px 0; border:1px solid var(--line);
  border-radius:8px; -webkit-overflow-scrolling:touch; scrollbar-width:thin;
  scrollbar-color:#d2d2d7 #f5f5f7;}
.edge0-card table{border-collapse:collapse; width:100%; font-size:13px; margin:0;}
.edge0-card th,.edge0-card td{border-bottom:1px solid #e5e5e7; padding:12px 13px; text-align:left; min-width:68px;}
.edge0-card th{background:#f5f5f7; color:#6e6e73; font-weight:500; white-space:nowrap; font-size:11px;}
.edge0-card tr:last-child td{border-bottom:0;}
.edge0-card tr:nth-child(even) td{background:#fafafa;}
.edge0-card tr.current-model td,.edge0-card tr.current-build td{background:#f5f5f7;}
.edge0-card tr.current-model td:first-child,.edge0-card tr.current-build td:first-child{box-shadow:inset 3px 0 #d2d2d7;}
.edge0-card tr.current-model td:first-child{color:var(--ink);}
.edge0-card td:first-child{min-width:155px;}
.edge0-card .speed-table-wrap,.edge0-card .speed-legend{--speed-header:var(--signal);
  --speed-highlight:#D8EEEF;}
.edge0-card .speed-table-wrap{border:0; scrollbar-color:var(--steel) var(--ivory);}
.edge0-card .speed-table{color:var(--navy);}
.edge0-card .speed-table th,.edge0-card .speed-table td{border:0;}
.edge0-card .speed-table th{color:var(--ivory); background:var(--speed-header);
  font-weight:700; padding-top:7px; padding-bottom:7px;}
.edge0-card .speed-table tbody tr td{background:#fff;}
.edge0-card .speed-table tbody tr.current-build td{background:var(--speed-highlight);}
.edge0-card .speed-table tr.current-build td:first-child{box-shadow:none;}
.edge0-card .speed-table tr.current-build td:nth-child(2){color:var(--navy); font-weight:600;}
.edge0-card .speed-table strong{color:var(--navy);}
.edge0-card .speed-legend{display:inline-block; color:var(--steel); font-size:12px; line-height:1.7;}
.edge0-card .speed-legend::before{content:""; display:inline-block; width:8px; height:8px;
  margin-right:8px; border-radius:2px; background:var(--ice);}
.edge0-card .speed-legend strong{color:var(--navy); font-weight:500;}
.edge0-card blockquote{margin:22px 0; padding:4px 20px; border-left:2px solid #d2d2d7;
  background:#fafafa; color:#6e6e73;}
.edge0-card blockquote p{margin:10px 0;}
.edge0-card hr{border:none; border-top:1px solid var(--line); margin:38px 0;}
.edge0-card ul,.edge0-card ol{padding-left:23px; margin:16px 0;}
.edge0-card li{margin:8px 0; padding-left:2px;}
.edge0-card li::marker{color:#86868b;}
.edge0-card details{margin:24px 0; padding:0 0 0; border-top:1px solid var(--line);
  border-bottom:1px solid var(--line);}
.edge0-card details + details{border-top:0; margin-top:-24px;}
.edge0-card summary{color:#515154; font-weight:500; font-size:14px; line-height:1.6;
  padding:17px 4px; cursor:pointer; list-style:none; display:flex; align-items:center; gap:12px;}
.edge0-card summary::-webkit-details-marker{display:none;}
.edge0-card summary::before{content:""; width:6px; height:6px; border-right:1.5px solid #86868b;
  border-bottom:1.5px solid #86868b; transform:rotate(-45deg); flex-shrink:0; margin:0 3px;
  transition:transform .16s ease;}
.edge0-card details[open] > summary::before{transform:rotate(45deg) translate(-1px,-1px);}
.edge0-card summary:hover{color:var(--ink);}
.edge0-card details[open]{padding-bottom:20px;}
.edge0-card details[open] > summary{margin-bottom:2px;}
.edge0-card details > :not(summary){margin-left:3px; margin-right:3px;}
.edge0-card details > :last-child{margin-bottom:0;}
.edge0-card details .table-scroll{margin-top:13px;}
.edge0-card details p{font-size:14px; color:var(--muted);}
.edge0-card .demo-section{margin:58px 0 0; border:0; scroll-margin-top:24px;}
.edge0-card .demo-section > summary{padding:0 0 13px; border-bottom:1px solid var(--line);
  display:grid; grid-template-columns:8px minmax(0,1fr); column-gap:14px; row-gap:7px;
  font-size:25px; font-weight:600; letter-spacing:-.5px; line-height:1.35; color:var(--ink);}
.edge0-card .demo-section > summary > .demo-title{grid-column:2; grid-row:1; margin:0; padding:0;
  border:0; font:inherit; color:inherit; letter-spacing:inherit;}
.edge0-card .demo-section > summary > .demo-model-note{grid-column:2; grid-row:2;
  font-size:12px; font-weight:400; line-height:1.65; letter-spacing:0; color:var(--steel);}
.edge0-card .demo-section[open]{padding-bottom:0;}
.edge0-card .demo-section[open] > summary{margin-bottom:0;}
.edge0-card .demo-section > summary::before{grid-column:1; grid-row:1; border-color:var(--steel);}
.edge0-card .showcase-player{max-width:480px; margin:32px auto 48px;}
.edge0-card .showcase-screen{position:relative;}
.edge0-card .showcase-video{display:block; width:100%; height:auto; aspect-ratio:16/9; object-fit:contain;
  border-radius:12px; background:#000; box-shadow:0 8px 32px #1D35570C;}
.edge0-card .showcase-play{position:absolute; left:50%; top:50%; transform:translate(-50%,-50%);
  display:grid; place-items:center; width:54px; height:54px; border:1px solid #FFFFFFAA;
  border-radius:50%; background:linear-gradient(135deg,#8EE4CF,#55BFC4 34%,#ADA0D4 66%,#F1B5CC);
  box-shadow:inset 0 2px 7px #FFFFFF55,0 4px 18px #1D355733; cursor:pointer;}
.edge0-card .showcase-play::after{content:""; display:block; margin-left:4px; width:0; height:0;
  border-top:9px solid transparent; border-bottom:9px solid transparent; border-left:15px solid white;}
.edge0-card .showcase-play[hidden]{display:none;}
.edge0-card .showcase-play:focus-visible{outline:3px solid var(--steel); outline-offset:4px;}
.edge0-card .showcase-meta{display:flex; gap:16px; align-items:center; justify-content:space-between;
  margin:14px 0 22px; color:var(--steel); font-size:11px; line-height:1.5;}
.edge0-card .showcase-meta small{font-size:inherit;}
.edge0-card .showcase-direct{color:var(--steel); font-size:11px; text-decoration:none;}
.edge0-card .showcase-seek-head{display:flex; justify-content:space-between; align-items:baseline; gap:16px;
  min-height:20px; color:var(--navy); font-size:12px;}
.edge0-card .showcase-current{font-weight:500;}
.edge0-card .showcase-time{flex:none; color:var(--steel); font-size:11px; font-variant-numeric:tabular-nums;}
.edge0-card .showcase-seek{position:relative; height:40px; margin:0 8px;}
.edge0-card .showcase-segments{position:absolute; inset:17px 0 auto; height:6px; display:flex; gap:3px; pointer-events:none;}
.edge0-card .showcase-segments span{height:6px; border-radius:4px; background:linear-gradient(to right,
  var(--steel) var(--chapter-fill,0%),var(--ice) var(--chapter-fill,0%));}
.edge0-card .showcase-slider{-webkit-appearance:none; appearance:none; display:block; position:absolute; inset:0;
  width:calc(100% + 16px); height:40px; margin:0 -8px; padding:0; background:transparent; cursor:pointer;}
.edge0-card .showcase-slider::-webkit-slider-runnable-track{height:6px; background:transparent;}
.edge0-card .showcase-slider::-moz-range-track{height:6px; background:transparent;}
.edge0-card .showcase-slider::-webkit-slider-thumb{-webkit-appearance:none; width:16px; height:16px; margin-top:-5px;
  border:2px solid white; border-radius:50%; background:var(--signal); box-shadow:0 0 0 1px #E6394655,0 2px 6px #1D355722;}
.edge0-card .showcase-slider::-moz-range-thumb{width:12px; height:12px; border:2px solid white; border-radius:50%;
  background:var(--signal); box-shadow:0 0 0 1px #E6394655;}
.edge0-card .showcase-slider:focus-visible{outline:2px solid var(--steel); outline-offset:2px; border-radius:6px;}
.edge0-card .demo-chapters{display:grid; grid-template-columns:repeat(5,minmax(0,1fr)); gap:10px; margin-top:3px;}
.edge0-card .demo-chapters a{min-width:0; display:flex; flex-direction:column; align-items:flex-start; gap:5px;
  padding:10px 8px; border-radius:7px; color:var(--steel); font-size:11px; line-height:1.45; text-decoration:none;
  transition:background .15s ease,color .15s ease;}
.edge0-card .demo-chapters small{font-size:10px; letter-spacing:.3px; font-variant-numeric:tabular-nums;}
.edge0-card .demo-chapters a[aria-current="true"]{background:#A8DADC33; color:var(--navy);}
.edge0-card .demo-chapters a:hover{background:#A8DADC26; color:var(--navy);}
.edge0-card .demo-chapters a:focus-visible{outline:2px solid var(--steel); outline-offset:2px;}
.edge0-card{width:100%;max-width:100%;min-width:0;text-align:left;}
.edge0-card p{font-size:15px;}
.edge0-card table{display:table;width:100%;border-collapse:collapse!important;border-spacing:0;border:0;border-radius:0;overflow:visible;}
.edge0-card .speed-table{font-size:13px;}
.edge0-card .speed-table,.edge0-card .speed-table thead,.edge0-card .speed-table tbody,.edge0-card .speed-table tr,.edge0-card .speed-table th,.edge0-card .speed-table td{border:0!important;}
.edge0-card .badge-github a,.edge0-card .badge-models a{display:inline-block!important;}
.edge0-card .card-hero .badge{margin:3px;height:20px;width:auto;vertical-align:middle;}
.edge0-card .demo-section>summary>.demo-title{margin:0;padding:0;border:0;font:inherit;letter-spacing:inherit;color:inherit;grid-column:2;grid-row:1;}
.edge0-card .showcase-screen{max-width:480px;margin:0 auto;}
.edge0-card .showcase-screen video{width:100%;max-width:480px;height:auto;margin:0;}
.edge0-card .showcase-screen .video-container{width:100%;max-width:480px;margin:0;}
.edge0-card .demo-chapters{grid-template-columns:repeat(auto-fit,minmax(76px,1fr));}
</style>

<div class="edge0-card model-card">
<div align="center" class="card-hero">
<p><img alt="Edge0" class="hero-image" src="assets/hero.png" width="100%"/></p>
<h1 id="edge0-145b-nvfp4">Edge0-145B · NVFP4</h1>
<div align="center" class="editorial-art">
<p><img alt="Think big. Run here. A 145B agent. On one consumer GPU. Sparse. By design. Latent MoE. Trained natively at 4 bits." class="editorial-image" src="assets/editorial-pair.svg" width="720"/></p>
</div>
<p align="center" class="badge-github"><a href="https://github.com/Edge0-AI/Edge0"><img alt="GitHub" class="badge" height="20" src="assets/badges/github-badge.svg"/></a> <a href="https://github.com/Edge0-AI/Edge0/releases"><img alt="Windows releases" class="badge" height="20" src="assets/badges/windows-releases-badge.svg"/></a></p>
<p align="center" class="badge-models"><a href="https://huggingface.co/Edge0/Edge0-145B-q3k"><img alt="Edge0-145B-q3k" class="badge" height="20" src="assets/badges/Edge0-145B-q3k-badge.svg"/></a> <a href="https://huggingface.co/Edge0/Edge0-120B-q3k"><img alt="Edge0-120B-q3k" class="badge" height="20" src="assets/badges/Edge0-120B-q3k-badge.svg"/></a> <a href="https://huggingface.co/Edge0/Edge0-120B-NVFP4"><img alt="Edge0-120B-NVFP4" class="badge" height="20" src="assets/badges/Edge0-120B-NVFP4-badge.svg"/></a></p>
<p class="section-nav"><a href="#quality">Quality</a> · <a href="#speed">Speed</a> · <a href="#architecture">Architecture</a> · <a href="#quick-start">Quick start</a></p>
</div>
<hr/>
<div class="model-context">
<p><strong>A 145B agent, running on one consumer GPU.</strong></p>
<p><strong>98.3 tok/s agent decode · 3165.7 tok/s 16K prefill · 86.5 terminal-bench 2.1.</strong></p>
<p>Throughput measured on <strong>RTX 5090 D 24 GB · NVFP4</strong>.</p>
</div>
<p>A latent sparse MoE: 48 layers, 512 experts, 10 active per token, 256K context, trained natively at 4 bits.</p>
<div class="repo-summary">
<p><strong>This repository:</strong> Edge0-145B · <strong>NVFP4 </strong> </p>
<p><small class="repo-note">ⓘ With residency enabled, <strong>145B NVFP4</strong> used <strong>23.25 GiB VRAM + 38.27 GiB host RAM</strong> on RTX 5090 D 24 GB. </small></p>
</div>
<details open="">
<summary><b>Model overview</b></summary>
<p><strong>NVFP4 · RTX 5090 D 24 GB.</strong> Prefill throughput uses a 16K-token input; weight sizes are listed separately for each format.</p>
<p><a href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/model-overview.svg" title="Open full-size model overview"><img alt="Edge0-145B NVFP4 model overview: architecture, benchmark scores, throughput and weight sizes" src="assets/model-overview.svg" width="100%"/></a></p>
</details>
<h2 id="quality">Quality</h2>
<p>Higher is better. Results are listed by model.</p>
<p><a class="quality-figure-link" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/quality/quality-overview.svg" title="Open full-size chart"><img alt="Six-model Quality comparison: DeepSWE 1.1, Terminal-Bench 2.1, HLE, GPQA-Diamond, IFBench, AA-LCR. All panels use a 0–100 scale. Full model names, scores and missing values are shown in the chart." class="quality-chart" src="assets/quality/quality-overview.svg" width="1400"/></a></p>
<h2 id="speed">Speed</h2>
<p>Throughput is shown in <strong>tok/s</strong>, grouped by hardware and build. The summary above uses <strong>NVFP4</strong> on RTX 5090 D 24 GB.</p>
<details open>
<summary><b>Full speed table</b> · all reported machines and builds</summary>
<p><small class="speed-legend">Highlighted rows show the <strong>NVFP4</strong> build in this repository.</small></p>
<div aria-label="Model card table; scroll horizontally to see all columns" class="table-scroll speed-table-wrap" role="region" tabindex="0"><table class="speed-table">
<thead>
<tr>
<th>Machine</th>
<th>Build</th>
<th style="text-align: right;">4K prefill</th>
<th style="text-align: right;">16K prefill</th>
<th style="text-align: right;">Chat decode</th>
<th style="text-align: right;">Agent decode</th>
</tr>
</thead>
<tbody>
<tr>
<td>RTX 5090 D 24 GB</td>
<td>Q3_K</td>
<td style="text-align: right;">2169.9</td>
<td style="text-align: right;">3554.9</td>
<td style="text-align: right;">99.7</td>
<td style="text-align: right;"><strong>121.2</strong></td>
</tr>
<tr class="current-build">
<td>RTX 5090 D 24 GB</td>
<td>NVFP4</td>
<td style="text-align: right;">1839.6</td>
<td style="text-align: right;">3165.7</td>
<td style="text-align: right;">79.6</td>
<td style="text-align: right;">98.3</td>
</tr>
<tr>
<td>RTX 4090 24 GB</td>
<td>Q3_K</td>
<td style="text-align: right;">2042.0</td>
<td style="text-align: right;">3179.4</td>
<td style="text-align: right;">69.9</td>
<td style="text-align: right;">91.7</td>
</tr>
<tr>
<td>RTX Spark N1X 60 GiB</td>
<td>Q3_K</td>
<td style="text-align: right;">674.0</td>
<td style="text-align: right;">770.3</td>
<td style="text-align: right;">37.7</td>
<td style="text-align: right;">36.7</td>
</tr>
<tr class="current-build">
<td>RTX Spark N1X 60 GiB</td>
<td>NVFP4</td>
<td style="text-align: right;">821.2</td>
<td style="text-align: right;">922.4</td>
<td style="text-align: right;">40.1</td>
<td style="text-align: right;">46.0</td>
</tr>
<tr>
<td>Apple M2 Ultra (Metal)</td>
<td>Q3_K</td>
<td style="text-align: right;">755.6</td>
<td style="text-align: right;">800.3</td>
<td style="text-align: right;">60.9</td>
<td style="text-align: right;">57.0</td>
</tr>
<tr>
<td>16 GiB VRAM + 20 GiB RAM</td>
<td>Q3_K</td>
<td style="text-align: right;">249</td>
<td style="text-align: right;">335</td>
<td style="text-align: right;">36.8</td>
<td style="text-align: right;">44.9</td>
</tr>
<tr>
<td>12 GiB VRAM + 20 GiB RAM</td>
<td>Q3_K</td>
<td style="text-align: right;">221</td>
<td style="text-align: right;">222</td>
<td style="text-align: right;">22.4</td>
<td style="text-align: right;">23.5</td>
</tr>
</tbody>
</table></div>
</details>
<details open>
<summary><b>Measurement conditions</b> · constrained budget and memory residency</summary>
<p>The <code>16 GiB VRAM + 20 GiB RAM</code> rows are the same RTX 5090 D 24 GB, not a different card. VRAM is capped with <code>--vram-reserve-mib 8192</code>. System RAM is capped by a cgroup (<code>MemoryMax=20G</code>, no swap). With residency on, the same card holds <strong>Q3_K</strong> (40.81 GiB of weights)
as 23.14 GiB VRAM + 31.28 GiB RAM, and <strong>NVFP4</strong> (47.79 GiB) as 23.25 GiB VRAM + 38.27 GiB RAM — disjoint expert sets,
nothing counted twice.</p>
<p>The <code>12 GiB VRAM + 20 GiB RAM</code> rows tighten the same RTX 5090 D 24 GB to a 12 GiB engine VRAM budget (<code>--vram-reserve-mib 12288</code>) under the same 20 GiB cgroup (<code>MemoryMax=20G</code>, no swap). They run the <strong>Q3_K</strong> build with the Q8 dense table, no MTP draft head (<code>--spec 4 --suffix-draft 3</code>), <code>--prefill 512</code>, and <code>--mmap-experts --resident-budget-gib 20</code> (clamped to &asymp;14 GiB by the cgroup); the remaining experts stream from SSD. These are test limits, not minimum system requirements. See <a href="#usage-help">Usage &amp; troubleshooting</a> for the full small-card setup.</p>
<p class="section-nav">See <a href="#usage-help">Usage &amp; troubleshooting</a> for memory-saving options.</p>
</details>
<details class="demo-section" id="in-action" open="">
<summary><h2 class="demo-title">In action</h2></summary>
<div class="showcase-player" data-duration="121.301043">
<div class="showcase-screen">
<video aria-label="Edge0-145B demonstrations" class="showcase-video" controls="" playsinline="" poster="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase-frame.jpg" preload="metadata" width="480">
<source src="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4" type="video/mp4"/>
<track kind="chapters" label="Chapters" src="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase-chapters.vtt" srclang="en"/>
<a href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/blob/main/assets/demos/showcase.mp4">Watch the Edge0 demonstrations</a>
</video>
</div>
<p class="showcase-meta"><small>Recorded with Edge0-145B</small><a class="showcase-direct" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/blob/main/assets/demos/showcase.mp4">Open video ↗</a></p>
<nav aria-label="Video chapters" class="demo-chapters">
<a aria-label="Local Chat &amp; Coding, 0:00" data-end="36.56666666666667" data-start="0.0" data-title="Local Chat &amp; Coding" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4#t=0.0"><small>0:00</small><span>Chat &amp; Coding</span></a>
<a aria-label="Video Production, 0:36" data-end="58.4" data-start="36.56666666666667" data-title="Video Production" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4#t=36.56666666666667"><small>0:36</small><span>Video Production</span></a>
<a aria-label="Manhattan in 3D, 0:58" data-end="79.56666666666666" data-start="58.4" data-title="Manhattan in 3D" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4#t=58.4"><small>0:58</small><span>Manhattan in 3D</span></a>
<a aria-label="Cathedral World Building, 1:19" data-end="98.93333333333334" data-start="79.56666666666666" data-title="Cathedral World Building" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4#t=79.56666666666666"><small>1:19</small><span>3D Cathedral</span></a>
<a aria-label="PCB Design, 1:38" data-end="121.301043" data-start="98.93333333333334" data-title="PCB Design" href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/demos/showcase.mp4#t=98.93333333333334"><small>1:38</small><span>PCB Design</span></a>
</nav>
</div>
</details>
<h2 id="architecture">Architecture</h2>
<details open>
<summary><b>Architecture details</b> · diagram, configuration and expert routing</summary>
<p><a href="https://huggingface.co/Edge0/Edge0-145B-NVFP4/resolve/main/assets/arch.svg" title="Open full-size architecture diagram"><img alt="Edge0-145B architecture and expert routing" src="assets/arch.svg" width="100%"/></a></p>
<div aria-label="Model card table; scroll horizontally to see all columns" class="table-scroll" role="region" tabindex="0"><table>
<thead>
<tr>
<th></th>
<th>Edge0-145B</th>
</tr>
</thead>
<tbody>
<tr>
<td>Total parameters</td>
<td><strong>145B</strong> · <strong>10.12 B active per token</strong></td>
</tr>
<tr>
<td>Layers</td>
<td>48 · hybrid attention, full every 4 layers + sparse indexer</td>
</tr>
<tr>
<td>Experts</td>
<td><strong>512 per layer</strong>, 10 routed + 1 shared active per token, SwiGLU width 640</td>
</tr>
<tr>
<td>Expert pool</td>
<td>84.6 B parameters at latent width ℓ = <strong>1792</strong></td>
</tr>
<tr>
<td>PLE</td>
<td><strong>51.2 B</strong> n-gram embedding table, fp8, memory-mapped</td>
</tr>
<tr>
<td>MTP head</td>
<td>1.26 B draft model, 512 experts, for speculative decode</td>
</tr>
<tr>
<td>Context</td>
<td>256K · KV int8 with 32K resident</td>
</tr>
<tr>
<td>Builds</td>
<td>Q3_K 40.81 GiB · NVFP4 47.79 GiB</td>
</tr>
</tbody>
</table></div>
<p><strong>Left</strong>: the stack the model actually runs — token embedding, 48 hybrid-attention layers, final norm, lm_head.
<strong>Right</strong>: one layer opened up. Attention reads and writes the full 2560 width. <strong>D</strong> then projects the token down to
ℓ per routing group, the router picks 10 of 512 experts, each expert runs ℓ → 640 → ℓ, a shared expert runs
alongside, and <strong>U</strong> projects the combined result back to 2560.</p>
<p>Everything outside the expert pool stays at 2560: D and U are shared, one copy per layer. Only the 512 experts work at
ℓ. That is why an expert is a fixed <strong>1,630,720 B</strong> record — the engine keeps a few thousand of them on the card and streams
the rest from host memory, at a cost that tracks how many experts are resident and not how big the model is.</p>
</details>

## Rebuilt, not converted

Edge0 is not a relabelled Qwen3.8-Flash-Next, and not a re-quantization of it: we rebuilt the model's
expert layer. In all 48 layers the routed MoE experts were replaced by a latent MoE — the token is
projected down to a narrow latent width (ℓ = 1792 on the 145B, 1280 on the 120B), the experts run
there, and the result is projected back to 2560. One expert is now a fixed 1.63 MB (145B) or
1.16 MB (120B) record instead of a full-width block. The architecture itself changed, so the model was
post-trained afterwards on a long schedule over our own data; it is trained natively at 4 bits, not
quantized after the fact.

That is what matters for a device. On-device inference runs out of memory capacity and memory bandwidth
long before it runs out of compute, and the expert bytes a token pushes through set both decode speed
and how much memory a machine needs:

| | Qwen3.8-Flash-Next | Edge0-145B | Edge0-120B |
|---|---:|---:|---:|
| Routed expert parameters | 120.8 B | **84.6 B** (−30%) | **60.4 B** (−50%) |
| Expert parameters per token | 2.36 B | **1.65 B** | **1.18 B** |
| Expert pool on disk · NVFP4 builds | 63.2 GiB | **44.3 GiB** | **31.6 GiB** |
| Decode · 16 GiB VRAM + 20 GiB RAM | 18.5 tok/s | **36.8 tok/s** | **62.4 tok/s** |

On that budget, decode runs 2.0× Qwen's on the 145B and 3.4× on the 120B. Quality holds through the
rebuild: the 145B is ahead on the agentic benchmarks — deepswe 1.1 64.6 vs 58.7, terminal-bench 2.1
86.5 vs 83.3 — and level elsewhere.

<h2 id="quick-start">Quick start</h2>
<p>Each model folder (Q3_K or NVFP4, Edge0-145B or Edge0-120B) runs on <a href="https://github.com/Edge0-AI/Edge0">Edge0</a>
through <code>edge0-infer</code>, from the command line or from the Edge0 Chat desktop app. Keep the folder as it is:
<code>edge0-infer</code> identifies the model by <code>model_name</code> in its <code>model.json</code>.</p>
<details open>
<summary><b>Platform &amp; documentation notes</b></summary>
<p>The build instructions cover NVIDIA GPUs with CUDA 13 on Linux or Windows (x64 or ARM64), and Metal on Apple Silicon macOS. Both the Q3_K and the NVFP4 folders run on either; on Metal the NVFP4 expert pool stays in the machine's unified memory.</p>
</details>
<details open="">
<summary><b>Build &amp; command-line instructions</b></summary>
<p><strong>1. Build.</strong> Clone the repository, then build two parts:</p>
<pre><code class="language-bash">git clone https://github.com/Edge0-AI/Edge0.git edge0
</code></pre>
<ul>
<li><strong>edge0-strata</strong>, the engine: follow "Build (Linux, NVIDIA)", "Build (Windows, NVIDIA)" or "Build (macOS, Apple Silicon)" in <a href="https://github.com/Edge0-AI/Edge0/blob/main/edge0-strata/EDGE0.md">edge0-strata/EDGE0.md</a>.</li>
<li><strong>edge0-infer</strong>: follow "Build" (Linux) or "Windows" in <a href="https://github.com/Edge0-AI/Edge0/blob/main/edge0-infer/docs/getting-started.md">edge0-infer/docs/getting-started.md</a>.</li>
</ul>
<p><strong>2. Run from the command line.</strong> Pass one model folder, such as <code>Edge0-145B-NVFP4</code>, not the folder that holds several of
them. On Windows the program is <code>build\windows\bin\edge0-infer.exe</code>.</p>
<pre><code class="language-bash">build/cpu/bin/edge0-infer serve --model /path/to/Edge0-145B-NVFP4
</code></pre>
<p>It prints <code>Listening: http://127.0.0.1:8000 model=&lt;name&gt;</code> once the weights are loaded, where <code>&lt;name&gt;</code> is <code>edge0_145b</code>,
and then serves OpenAI- and Anthropic-compatible endpoints:</p>
<pre><code class="language-bash">curl localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{"model":"edge0_145b","messages":[{"role":"user","content":"hi"}],"max_tokens":64}'
</code></pre>
<p><code>edge0-infer chat --model /path/to/Edge0-145B-NVFP4</code> opens an interactive session instead. More options:
<a href="https://github.com/Edge0-AI/Edge0/blob/main/edge0-infer/docs/edge0-145b.md">edge0-infer/docs/edge0-145b.md</a>.</p>
</details>
<details open>
<summary><b>Desktop app</b> · Windows</summary>
<p>⬇️ <strong>Download:</strong> <strong><a href="https://github.com/Edge0-AI/Edge0/releases">GitHub Releases</a></strong></p>
<p>Open <strong>Edge0 Chat</strong> on Windows and choose a model folder, such as <code>Edge0-145B-NVFP4</code>, to run the model locally. To build the app from source, follow the <a href="https://github.com/Edge0-AI/Edge0/blob/main/edge0-infer/examples/desktop/README.md">Edge0 Chat</a> guide.</p>
</details>
<details open>
<summary><b>Images</b> · the vision encoder in this folder</summary>
<p><code>mmproj-BF16.gguf</code> is the vision encoder: it turns a picture into the embedding rows the language model
reads, and the language model itself is unchanged. Reading pictures also needs a small helper program,
<code>strata-vision</code>. It comes from one of these two places:</p>
<ul>
<li><strong>The prebuilt engine archive.</strong> <code>setup.py</code> and <code>strata launch</code> download one and
unpack <code>strata</code> and <code>strata-vision</code> side by side, so an install that ran with <code>--vision</code>
already has the helper next to the engine.</li>
</ul>
<p>Or build its source, which is in this repository at <code>edge0-strata/tools/vision/</code>. It is a separate CMake
project that the main build does not compile, and it needs a llama.cpp checkout at the engine's pinned commit
(<code>third_party/llama.cpp</code> is not vendored; a tree whose main CMake fetched it has one under
<code>build/_deps/strata_llamacpp-src</code>):</p>

<pre><code class="language-sh">cmake -S edge0-strata/tools/vision -B build-vision -DLLAMA_DIR=/path/to/llama.cpp
cmake --build build-vision --target strata-vision</code></pre>

<p>Put the result next to <code>strata</code>.</p>
<p>Then two entries in <code>model.json</code>, next to <code>args</code>:</p>
<pre><code class="language-json">"exe": "/path/to/strata",
"vision": {
  "exe": "/path/to/strata-vision",
  "mmproj": "./mmproj-BF16.gguf",
  "model": "./model-skeleton.gguf",
  "gpu": false,
  "max_tokens": 1024
}
</code></pre>
<ul>
<li><code>"model"</code> is the text backbone this folder already uses.</li>
<li><code>"gpu": false</code> runs the encoder on the CPU and takes no VRAM. <code>true</code> runs it on the card and
keeps about 1.4 GB free for it, so the expert cache gets that much less.</li>
<li><code>"max_tokens"</code> caps how many tokens one picture becomes. More keeps more detail and takes longer.</li>
</ul>
<p>Add <code>--vision</code> to <code>"args"</code> as well. The server adds it by itself when the config has a
<code>vision</code> section and prints a line saying so; writing it keeps the two in step. Without the helper beside the
engine, a request that carries a picture fails with <code>this server was started without the vision encoder</code>.</p>
<p>The log says <code>images on</code> once the encoder is up. Then send a picture: <code>/image /path/to.png</code>
followed by the question in the terminal chat, or an <code>image_url</code> part over the OpenAI API. That part takes a
<code>data:</code> URL, an <code>http(s)</code> URL or a local file path.</p>
<pre><code class="language-bash">curl localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{
  "model": "edge0_145b",
  "messages": [{"role": "user", "content": [
    {"type": "text", "text": "What does this picture show?"},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,&lt;base64&gt;"}}
  ]}],
  "max_tokens": 256
}'
</code></pre>
<p>Add <code>"reasoning_effort": "none"</code> to get the answer without a thinking pass first; otherwise the reply can
come back empty with the token budget spent on the reasoning.</p>
</details>
<p><a name="memory-update"></a>
<a name="deployment-qa"></a>
<a name="memory-deployment"></a>
<a name="usage-help"></a></p>
<h2 id="usage-troubleshooting">Usage &amp; troubleshooting</h2>
<p>145B / 120B · Memory options</p>
<details>
<summary><a name="low-memory"></a>Low RAM · SSD offload</summary>
<ul>
<li><strong>SSD mode:</strong> <code>--mmap-experts</code> — read uncached experts from SSD on demand.</li>
<li><strong>Low-memory arguments:</strong></li>
</ul>
<pre><code class="language-bash">--resident-budget-gib 20 --expert-profile expert-profile.bin
</code></pre>
<p>Up to <strong>20 GiB</strong> of host RAM for the expert cache; uncached experts are read from SSD on demand. Includes <code>--mmap-experts</code> and disjoint GPU/RAM caches automatically. The budget caps the <strong>expert cache only</strong> &mdash; not total process RAM &mdash; and is lowered further to fit available memory; the 20 GiB total-RAM test budget is a separate measurement condition.</p>
<p><strong>Required:</strong> a static expert profile. Replace <code>expert-profile.bin</code> with your profile path.</p>
</details>
<details>
<summary><a name="expert-residency"></a>GPU / RAM · avoid duplicate experts</summary>
<ul>
<li><strong>Enable:</strong> <code>--resident-experts</code> or <code>--resident-budget-gib N</code>, with a static <code>--expert-profile</code>.</li>
<li><strong>Effect:</strong> experts cached in VRAM need no duplicate RAM copy. Compatible with SSD loading.</li>
<li><strong>Default:</strong> off, prioritizing speed.</li>
<li><strong>Automatically enabled:</strong> <code>--mmap-experts</code>, <code>--resident-cpu-experts</code>, <code>--resident-pin</code>.</li>
</ul>
</details>
<details>
<summary><a name="small-card-12g"></a>12 GiB VRAM + 20 GiB RAM &middot; running on a small card</summary>
<p>If your machine only has a <strong>12 GiB VRAM</strong> card and <strong>20 GiB</strong> of system RAM, run the <strong>Q3_K</strong> build with the <strong>Q8</strong> dense table and these settings in the pack's <code>model.json</code> &mdash; the configuration behind the 12 GiB row above. Flags go in <code>args</code>:</p>
<pre><code class="language-bash">--vram-reserve-mib 700 --expert-cache auto
--mmap-experts --resident-budget-gib 20 --expert-profile expert-profile.bin
--conversation-cache-mib 0 --ple-io direct
--kv int8 --kv-resident 20480 --max-context 32768
--prefill 512 --spec 4 --suffix-draft 3
--pool-workers 24 --pcie-mode dma --pcie-frac 0.1
</code></pre>
<p>Environment variables in <code>env</code>:</p>
<pre><code class="language-bash">STRATA_LATENT_DENSE=./latent_dense_q8.bin
STRATA_LATENT_DENSE_FMT=8
STRATA_LATENT_A=./experts.bin
STRATA_LATENT_LABELS=./labels.bin
STRATA_EMBED_VRAM=1
STRATA_LATENT_FUSED=1
STRATA_LATENT_FUSED_DU=1
STRATA_LATENT_U_GPU=1
STRATA_LAT_FUSED_Q3K=1
</code></pre>
<ul>
<li><strong>VRAM reserve:</strong> <code>--vram-reserve-mib</code> is the VRAM you leave <em>free</em> for the desktop and driver, so the number depends on your card's total. On a 12 GiB card use <code>--vram-reserve-mib 700</code>; the engine sizes its caches from the remaining &asymp;11.3 GiB. (The speed row was measured by giving the engine a full 12 GiB budget on a 24 GiB card via <code>--vram-reserve-mib 12288</code>; a physical 12 GiB card leaves a little less for the engine, so the numbers are close approximations.)</li>
<li><strong>RAM:</strong> <code>--resident-budget-gib 20</code> caps the host-RAM expert cache; the engine clamps it to what is actually free (&asymp;14 GiB when 20 GiB is total) and streams the remaining experts from SSD. <code>--conversation-cache-mib 0</code> and <code>--ple-io direct</code> keep the conversation cache and the PLE table out of RAM. A static <code>--expert-profile</code> is required.</li>
<li><strong>No MTP head:</strong> leave out <code>--mtp</code>; <code>--spec 4 --suffix-draft 3</code> drafts from repeated context instead, saving &asymp;0.8 GiB VRAM.</li>
<li><strong>Prefill 512:</strong> keep <code>--prefill 512</code>. Chunks below 1024 let the idle CPU pool share the expert reads from SSD; at <code>--prefill 1024</code> a small card misses this and prefill runs 4&ndash;5&times; slower.</li>
</ul>
<p>Measured with this configuration (Q3_K, Q8 dense, 12 GiB engine VRAM budget, 20 GiB RAM): <strong>145B</strong> chat 22.4 / agent 23.5 tok/s, 4K 221 / 16K 222 tok/s; <strong>120B</strong> chat 35.3 / agent 43.5 tok/s, 4K 679 / 16K 734 tok/s.</p>
</details>

## Capacity controls · VRAM and host RAM

<a name="capacity-controls"></a>

All of these live in the pack's `model.json`, not on a command line: flags go in `args`, environment
variables in `env`, with paths relative to the pack directory. That is the same file the Edge0 Chat app
and `edge0-infer` read, so a change there applies to every launch.

```json
"args": ["--expert-cache", "auto", "--kv-resident", "32768", "--vram-reserve-mib", "1024"],
"env":  {"STRATA_LATENT_DENSE": "./latent_dense_q8.bin", "STRATA_LATENT_DENSE_FMT": "8",
         "STRATA_LAT_FUSED_Q3K": "0"}
```

### VRAM

| Flag | Meaning | Default |
|---|---|---|
| `--vram-reserve-mib N` | VRAM left free for the desktop, the driver and other processes; the engine sizes its caches from what remains. It means "leave this much free", not "give the engine this much", so the same number means different things on cards with different totals. Accepted but not listed by `--help`. | `700` |
| `--expert-cache N\|auto` | Expert slots kept in VRAM. `auto` derives the count from the free VRAM after the fixed footprint (skeleton projections ≈ 2.9 GiB, MTP draft ≈ 0.8 GiB, verify and draft-head buffers ≈ 0.45 GiB, plus the D/U device buffers). | `0` (none) |
| `--expert-cache-device1..3 N\|auto` | Use the second and third GPU as further expert-cache tiers. | off |
| `--kv-resident N` | Keep N KV cells per QSA layer in VRAM and stream the K/V from pinned host RAM, which frees VRAM for experts. | `0` (all in VRAM) |
| `--mtp-q4 all\|proj\|head` | 4-bit copies of the draft layer's projections and/or head, about 0.3–0.5 GiB less VRAM. | off |
| `--prefill CHUNK` | Prompt chunk size; bigger chunks need bigger device buffers. `auto` uses the largest chunk whose buffers the expert cache can lend. | `auto` |
| `--spec N` | The MTP drafter's verify window: how many tokens it proposes per check. A bigger window needs bigger verify buffers (about 180 MiB at 4) and pays off when every window has a fixed per-layer cost. The shipped packs use `4`; the macOS setup picks `2`, or `3` when the experts stream from SSD. Without an MTP layer there is no speculation. | `4` |

### Host RAM

| Flag | Meaning | Default |
|---|---|---|
| `--memory-limit-mib N` | The RAM this engine's container or cgroup may use, for when the container cannot see its own limit. | `0` (detect) |
| `--resident-budget-gib N` | Cap on the host-RAM expert cache. It holds only the experts the VRAM cache does **not** hold, so the two sets are disjoint, and the value is clamped to the RAM actually available minus headroom. Requires a static `--expert-profile`. | `0` (off) |
| `--resident-experts` | The same with the whole complement as the budget. | off |
| `--mmap-experts` | Read experts through the pack file on demand instead of pinning them, so nothing is resident in RAM. | off |
| `--conversation-cache-mib N` | Conversation cache. `0` turns it off. | adaptive |
| `--ple-io direct\|mmap\|ram` | The PLE table on its SSD: `direct` reads it unbuffered and it never enters memory; `ram` locks the whole table in RAM. | `direct` |
| `--kv-resident N` | The streamed K/V lives in pinned host RAM (1.55 GiB at 32768 cells). | — |

## Community builds

Third-party conversions of this model. They are **not made or endorsed by Edge0**; this model's licence and the
upstream terms still apply to them.

- **[infosave/Edge0-145B-cmf](https://huggingface.co/infosave/Edge0-145B-cmf)** — a single-file CMF build for
  [cortiq](https://github.com/infosave2007/cmf) (Rust, CUDA and Vulkan, OpenAI-compatible). The routed experts are this
  NVFP4 build's weights copied bit for bit; the D/U projections and the shared expert are re-encoded to 8 bits and the
  router to 16. The author reports it running from 12 GB cards.

<h2 id="acknowledgements">Acknowledgements</h2>
<details open>
<summary><b>Credits &amp; license</b></summary>
<p>We thank the teams behind <a href="https://huggingface.co/Qwen/Qwen3.8-Flash-Next">Qwen3.8-Flash-Next</a> and <a href="https://github.com/Niko1221/Strata">Strata</a> for sharing their work with the community. Edge0 extends the Strata engine with latent MoE support.</p>
<p><strong>Model license.</strong> Edge0-specific contributions are provided under the <a href="LICENSE">Edge0 Community Model License 1.0</a>, only to the extent that Edge0 owns or is authorized to license them. Qwen materials remain subject to the <a href="third-party/Qwen-LICENSE.txt">Qwen Community License 1.0</a>. Both sets of applicable terms must be followed; see <a href="NOTICE">NOTICE</a> and the <a href="release-scope.json">release scope</a>.</p>
<p>The Edge0 license permits personal use, research, modification, noncommercial sharing, internal use, and commercial use of generated outputs, subject to applicable upstream conditions. Selling or commercially bundling the model, operating a commercial model service, or offering an independent commercial coding or office assistant using it requires a separate written Edge0 license. See LICENSE for the precise scope and exceptions.</p>
<p><strong>Commercial licensing:</strong> <a href="mailto:info@edge0.ai">info@edge0.ai</a>. Edge0 permission does not waive Qwen's separate authorization or attribution requirements. Framework code retains its Apache-2.0 license; Strata retains its MIT license. Existing valid grants are not revoked.</p>
</details>
</div>
