mirror of
https://github.com/hanzoai/skills.git
synced 2026-11-19 00:47:36 +00:00
326 lines
28 KiB
HTML
326 lines
28 KiB
HTML
<!doctype html>
|
||
<html lang="en"><head>
|
||
<meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
|
||
<title>hip-0904</title>
|
||
<meta name="color-scheme" content="dark">
|
||
<style>
|
||
/* Committed to one visual world: a dark instrument panel. No light variant —
|
||
the dot field and the hairline rules only read on black, and a naive
|
||
inversion would wash both out. Every colour is painted explicitly so the page
|
||
never borrows a host background. */
|
||
@font-face{font-family:'Zen';src:url('/fonts/Zen-Variable.woff2') format('woff2');font-weight:100 900;font-style:normal;font-display:swap}
|
||
@font-face{font-family:'Zen';src:url('/fonts/Zen-Italic-Variable.woff2') format('woff2');font-weight:100 900;font-style:italic;font-display:swap}
|
||
@font-face{font-family:'Zen Mono';src:url('/fonts/ZenMono-Variable.woff2') format('woff2');font-weight:100 900;font-style:normal;font-display:swap}
|
||
@font-face{font-family:'Zen Mono';src:url('/fonts/ZenMono-Italic-Variable.woff2') format('woff2');font-weight:100 900;font-style:italic;font-display:swap}
|
||
@font-face{font-family:'Zen Pixel Square';src:url('/fonts/ZenPixel-Square.woff2') format('woff2');font-style:normal;font-display:swap}
|
||
|
||
:root{
|
||
--ink:#f4f4f2; --soft:#9c9c98; --faint:#6a6a66; --ground:#000; --panel:#0a0a0a;
|
||
--line:#232323; --line-hi:#3a3a3a; --accent:#f4f4f2;
|
||
--font-zen-sans:'Zen',-apple-system,BlinkMacSystemFont,"Segoe UI",Helvetica,Arial,sans-serif;
|
||
--font-zen-mono:'Zen Mono',ui-monospace,"SF Mono",SFMono-Regular,Menlo,Consolas,monospace;
|
||
--font-zen-pixel-square:'Zen Pixel Square',monospace;
|
||
--mono:var(--font-zen-mono);
|
||
--sans:var(--font-zen-sans);
|
||
/* Controls are pills and surfaces are soft, as on hanzo.ai. */
|
||
--r:999px; --r-soft:14px;
|
||
}
|
||
*{box-sizing:border-box}
|
||
html{-webkit-text-size-adjust:100%}
|
||
body{margin:0;background:var(--ground);color:var(--ink);font-family:var(--sans);
|
||
font-size:16px;line-height:1.65;-webkit-font-smoothing:antialiased}
|
||
.wrap{max-width:60rem;margin:0 auto;padding:0 1.75rem}
|
||
|
||
/* The dot field. A halftone made of one repeating radial-gradient, faded out
|
||
along the diagonal by a mask so it reads as a wave rather than wallpaper. */
|
||
/* No overflow clip here: the dot field is inset to the box already, and a clip
|
||
would cut off the copy-prompt menu. */
|
||
.hero{position:relative;border-bottom:1px solid var(--line)}
|
||
.hero::before{content:"";position:absolute;inset:0;
|
||
background-image:radial-gradient(var(--line-hi) 1px, transparent 1px);
|
||
background-size:7px 7px;
|
||
-webkit-mask-image:linear-gradient(115deg,#000 0%,rgba(0,0,0,.55) 38%,transparent 72%);
|
||
mask-image:linear-gradient(115deg,#000 0%,rgba(0,0,0,.55) 38%,transparent 72%);
|
||
pointer-events:none}
|
||
.hero .wrap{position:relative;padding-top:5.5rem;padding-bottom:4rem}
|
||
|
||
nav{border-bottom:1px solid var(--line)}
|
||
nav .in{max-width:60rem;margin:0 auto;padding:1.1rem 1.75rem;display:flex;
|
||
justify-content:space-between;align-items:center;gap:1rem}
|
||
.brand{font-family:var(--mono);font-size:.85rem;letter-spacing:.14em;
|
||
text-transform:uppercase;color:var(--ink);text-decoration:none;
|
||
display:inline-flex;align-items:center;gap:.6rem}
|
||
.brand i{font-style:normal;color:var(--faint)}
|
||
.mark{width:1.15em;height:1.15em;display:block;flex:0 0 auto}
|
||
nav a.up{font-family:var(--mono);font-size:.72rem;letter-spacing:.14em;
|
||
text-transform:uppercase;color:var(--soft);text-decoration:none}
|
||
nav a.up:hover{color:var(--ink)}
|
||
|
||
h1{font-family:var(--mono);font-size:clamp(2rem,5.4vw,3.4rem);line-height:1.12;
|
||
letter-spacing:-.02em;margin:0 0 1.1rem;font-weight:500;text-wrap:balance}
|
||
h1 em{font-style:normal;color:var(--faint)}
|
||
.lede{color:var(--soft);font-size:1.05rem;max-width:52ch;margin:0 0 2.25rem}
|
||
|
||
.btns{display:flex;flex-wrap:wrap;gap:.7rem;margin:0 0 .5rem;align-items:center}
|
||
.btn{display:inline-flex;align-items:center;gap:.5rem;font-family:var(--mono);
|
||
font-size:.72rem;letter-spacing:.14em;text-transform:uppercase;
|
||
padding:.8rem 1.4rem;border:1px solid var(--line-hi);border-radius:var(--r);
|
||
color:var(--ink);text-decoration:none;background:transparent;cursor:pointer;
|
||
line-height:1.2;font-weight:400}
|
||
.btn:hover{background:#141414;text-decoration:none;border-color:var(--soft)}
|
||
.btn.solid{background:var(--ink);color:#000;border-color:var(--ink)}
|
||
.btn.solid:hover{background:#d8d8d4}
|
||
.btn .chev{width:.7em;height:.7em;flex:0 0 auto;transition:transform .15s ease}
|
||
.btn[aria-expanded="true"] .chev{transform:rotate(180deg)}
|
||
|
||
/* Copy-prompt selector. Each row copies the prompt for its own target. */
|
||
.copy{position:relative;display:inline-flex}
|
||
.menu{position:absolute;top:calc(100% + .5rem);left:0;z-index:20;min-width:21rem;
|
||
background:var(--panel);border:1px solid var(--line-hi);border-radius:var(--r-soft);
|
||
padding:.4rem;display:flex;flex-direction:column;
|
||
box-shadow:0 12px 32px rgba(0,0,0,.75)}
|
||
.menu[hidden]{display:none}
|
||
.mi{display:flex;align-items:center;gap:.7rem;width:100%;text-align:left;
|
||
padding:.62rem .8rem;border:0;border-radius:calc(var(--r-soft) - .35rem);
|
||
background:transparent;color:var(--ink);font-family:var(--mono);font-size:.76rem;
|
||
letter-spacing:.04em;text-decoration:none;cursor:pointer;line-height:1.3;
|
||
white-space:nowrap}
|
||
.mi:hover,.mi:focus-visible{background:#181818;outline:none}
|
||
.mi .mark{width:1.05rem;height:1.05rem;color:var(--soft)}
|
||
.mi:hover .mark{color:var(--ink)}
|
||
.mi u{margin-left:auto;color:var(--faint);text-decoration:none;font-size:.68rem;
|
||
letter-spacing:.1em;text-transform:uppercase}
|
||
|
||
/* Works-with bar: one scrolling row, name and mark per agent. */
|
||
.strip{border-bottom:1px solid var(--line);background:var(--panel)}
|
||
.strip .in{max-width:60rem;margin:0 auto;padding:.9rem 1.75rem;display:flex;
|
||
align-items:center;gap:1.25rem;font-family:var(--mono);font-size:.72rem;
|
||
letter-spacing:.1em;text-transform:uppercase;color:var(--faint)}
|
||
.strip b{color:var(--soft);font-weight:400;flex:0 0 auto;
|
||
border-right:1px solid var(--line);padding-right:1.25rem;align-self:stretch;
|
||
display:flex;align-items:center}
|
||
.bar{display:flex;align-items:center;gap:.5rem;flex:1 1 auto;min-width:0;
|
||
overflow-x:auto;scrollbar-width:none;scroll-snap-type:x proximity;
|
||
-webkit-overflow-scrolling:touch;
|
||
-webkit-mask-image:linear-gradient(90deg,#000 calc(100% - 3rem),transparent);
|
||
mask-image:linear-gradient(90deg,#000 calc(100% - 3rem),transparent)}
|
||
.bar::-webkit-scrollbar{display:none}
|
||
.agent{display:inline-flex;align-items:center;gap:.45rem;flex:0 0 auto;
|
||
padding:.4rem .85rem;border:1px solid var(--line);border-radius:var(--r);
|
||
color:var(--soft);white-space:nowrap;scroll-snap-align:start;
|
||
text-decoration:none;background:var(--ground)}
|
||
.agent:hover{color:var(--ink);border-color:var(--line-hi);text-decoration:none}
|
||
.agent .mark{width:.95rem;height:.95rem}
|
||
|
||
h2{font-family:var(--mono);font-size:1.1rem;letter-spacing:-.01em;margin:2.75rem 0 .75rem;font-weight:500}
|
||
h3{font-size:1rem;margin:2rem 0 .5rem;font-weight:600}
|
||
h4,h5,h6{font-size:.94rem;margin:1.5rem 0 .5rem;font-weight:600}
|
||
p{margin:0 0 1rem;max-width:70ch}
|
||
ul,ol{margin:0 0 1rem;padding-left:1.35rem;max-width:70ch}
|
||
li{margin:.22rem 0}
|
||
a{color:var(--ink);text-decoration:underline;text-underline-offset:3px;
|
||
text-decoration-color:var(--line-hi)}
|
||
a:hover{text-decoration-color:var(--ink)}
|
||
a:focus-visible{outline:1px solid var(--ink);outline-offset:3px}
|
||
code{font-family:var(--mono);font-size:.85em;background:var(--panel);
|
||
border:1px solid var(--line);border-radius:5px;padding:.1em .36em}
|
||
pre{background:var(--panel);border:1px solid var(--line);padding:1rem 1.1rem;
|
||
border-radius:var(--r-soft);overflow-x:auto;margin:0 0 1.25rem}
|
||
pre code{background:none;border:0;padding:0;font-size:.82rem;line-height:1.7}
|
||
blockquote{margin:0 0 1rem;padding-left:1rem;border-left:1px solid var(--line-hi);color:var(--soft)}
|
||
hr{border:0;border-top:1px solid var(--line);margin:2.5rem 0}
|
||
table{border-collapse:collapse;width:100%;margin:0 0 1.25rem;font-size:.92rem;
|
||
display:block;overflow-x:auto}
|
||
th,td{border:1px solid var(--line);padding:.5rem .7rem;text-align:left}
|
||
|
||
.curl{display:block;background:var(--panel);border:1px solid var(--line);
|
||
border-radius:var(--r-soft);padding:.95rem 1.1rem;font-family:var(--mono);
|
||
font-size:.8rem;color:var(--soft);overflow-x:auto;margin:0 0 2.5rem;
|
||
white-space:pre;line-height:1.9}
|
||
.curl b{color:var(--ink);font-weight:400}
|
||
.curl i{font-style:normal;color:var(--faint)}
|
||
|
||
/* index */
|
||
#q{width:100%;padding:.85rem 1.2rem;font-family:var(--mono);font-size:.85rem;color:var(--ink);
|
||
background:var(--panel);border:1px solid var(--line);border-radius:var(--r);margin:0 0 .6rem}
|
||
#q::placeholder{color:var(--faint);text-transform:uppercase;letter-spacing:.1em;font-size:.72rem}
|
||
#q:focus{outline:none;border-color:var(--soft)}
|
||
#count{color:var(--faint);font-family:var(--mono);font-size:.72rem;letter-spacing:.1em;
|
||
text-transform:uppercase;margin:0 0 2rem}
|
||
.cat{margin:2.5rem 0 .5rem;font-family:var(--mono);font-size:.7rem;letter-spacing:.16em;
|
||
text-transform:uppercase;color:var(--faint);border-bottom:1px solid var(--line);
|
||
padding-bottom:.45rem}
|
||
.row{display:block;padding:.6rem .9rem;margin:0 -.9rem;border-bottom:1px solid var(--line);
|
||
border-radius:var(--r-soft);text-decoration:none}
|
||
.row:hover{background:#0a0a0a}
|
||
.nm{font-family:var(--mono);font-size:.84rem;color:var(--ink);display:block}
|
||
.ds{color:var(--soft);font-size:.9rem;display:block;margin-top:.12rem}
|
||
footer{margin-top:4rem;padding:2rem 0 4rem;border-top:1px solid var(--line);
|
||
color:var(--faint);font-size:.85rem}
|
||
footer a{color:var(--soft)}
|
||
.fmap{display:grid;gap:1.75rem;grid-template-columns:repeat(auto-fit,minmax(11rem,1fr));
|
||
margin:0 0 2rem;font-size:.85rem}
|
||
.fmap b{display:block;font-family:var(--mono);font-size:.68rem;letter-spacing:.16em;
|
||
text-transform:uppercase;color:var(--faint);font-weight:400;margin:0 0 .6rem;
|
||
border-bottom:1px solid var(--line);padding-bottom:.4rem}
|
||
.fmap a{display:inline-block;color:var(--ink);text-decoration:none;margin:.3rem .35rem 0 0}
|
||
.fmap a:hover{text-decoration:underline}
|
||
.fmap div{color:var(--faint);line-height:1.5}
|
||
.fend{border-top:1px solid var(--line);padding-top:1.25rem;font-size:.82rem}
|
||
|
||
@media (max-width:640px){
|
||
.hero .wrap{padding-top:3.25rem;padding-bottom:2.75rem}
|
||
.btns{gap:.5rem}
|
||
.btn{padding:.75rem 1.15rem;font-size:.68rem}
|
||
.copy,.copy .btn{width:100%;justify-content:center}
|
||
.menu{left:0;right:0;min-width:0}
|
||
.strip .in{flex-direction:column;align-items:flex-start;gap:.6rem;padding:.9rem 1.75rem}
|
||
.strip b{border-right:0;padding-right:0}
|
||
.bar{width:100%}
|
||
}
|
||
</style>
|
||
<link rel="alternate" type="text/markdown" href="/hip-0904.md">
|
||
|
||
<link rel="icon" href="/favicon.svg">
|
||
<link rel="canonical" href="https://hanzoskills.com/hip-0904">
|
||
</head>
|
||
<!-- Reading this with a machine? Every page here is also plain markdown.
|
||
This one: https://hanzoskills.com/hip-0904.md
|
||
Index: https://hanzoskills.com/llms.txt
|
||
Corpus: https://hanzoskills.com/all.md
|
||
Append .md to any path for the markdown twin. -->
|
||
<body>
|
||
<nav><div class="in">
|
||
<a class="brand" href="/"><svg class="mark" viewBox="0 0 67 67" fill="currentColor" aria-hidden="true"><path d="M22.21 67V44.6369H0V67H22.21Z"/><path d="M66.7038 22.3184H22.2534L0.0878906 44.6367H44.4634L66.7038 22.3184Z"/><path d="M22.21 0H0V22.3184H22.21V0Z"/><path d="M66.7198 0H44.5098V22.3184H66.7198V0Z"/><path d="M66.7198 67V44.6369H44.5098V67H66.7198Z"/></svg> HANZO<i>SKILLS</i></a>
|
||
<a class="up" href="/">All skills</a>
|
||
</div></nav>
|
||
|
||
<div class="wrap">
|
||
<h1>hip-0904</h1><p class="lede">HIP-904: Zen6 — Dense, Coder and Ternary Checkpoints. Status Final. Hanzo architectural specification.</p><h1>HIP-0904: Zen6 — Dense, Coder and Ternary Checkpoints</h1>
|
||
<h2>Abstract</h2>
|
||
<p>Zen6 is the current Zen generation (HIP-0039). It was published on 2026-09-21 as three open-weight Hugging Face repositories:</p>
|
||
<ul><li><strong><code>zenlm/zen6</code>.</strong> A 64-layer hybrid-attention model in mixed NVFP4 and FP8, with a DFlash 2</li></ul>
|
||
<p> drafter bundled beside it and YaRN scaling to 1,048,576 positions.</p>
|
||
<ul><li><strong><code>zenlm/zen6-coder</code>.</strong> A 512-expert mixture as GGUF IQ4_XS, with a one-layer MTP drafter.</li><li><strong><code>zenlm/zen6-flash</code>.</strong> A 64-layer model in ternary GGUF, with a vision projector and a</li></ul>
|
||
<p> DFlash 2 drafter.</p>
|
||
<p><strong>All three are re-hosts.</strong></p>
|
||
<ul><li>Every weight file is byte-identical to a named upstream file.</li><li>Only zen6 changes a model file: three rope fields in its <code>config.json</code>.</li><li>The only measured number in any of the three repositories is the upstream's GSM8K</li></ul>
|
||
<p> qualification of zen6.</p>
|
||
<p><code>zen6</code> and <code>zen6-coder</code> are also served SKU ids on <code>/v1/models</code>. Under HIP-0039 §2.3 they do not promise these weights.</p>
|
||
<p>This HIP records the layout of each repository, where each file came from, what has been measured and what is only claimed, and the defects found on 2026-09-25.</p>
|
||
<h2>Motivation</h2>
|
||
<p>Zen6 was published on Hugging Face on 2026-09-21 and is served on <code>/v1/models</code>. Until 2026-09-25, no HIP, catalog, site or doc named it.</p>
|
||
<p>Its cards also disagree with the files beside them. The zen6 card's architecture table gives a different head count, vocabulary and attention type from its <code>config.json</code>. The zen6-coder card describes a file format that is not in its repository.</p>
|
||
<p>A reader choosing between the three needs the artifacts, not the cards: what each file is, where it came from, what runs it, and which numbers carry a receipt.</p>
|
||
<h2>Specification</h2>
|
||
<p>The key words MUST, MUST NOT, SHOULD, SHOULD NOT and MAY are to be interpreted as in RFC 2119.</p>
|
||
<p>Every fact below was read on 2026-09-25 from the Hugging Face API (<code>https://huggingface.co/api/models/<id>?blobs=true</code>) and from the files at the revision given. GGUF metadata was read from the file headers with HTTP range requests.</p>
|
||
<h3>§1 Repositories</h3>
|
||
<p>| Checkpoint | Hugging Face @ revision | Bytes | Licence | Served SKU | |:--|:--|--:|:--|:--| | zen6 | <code>zenlm/zen6</code> @ <code>574c3066</code> | 25,794,104,008 | apache-2.0 | <code>zen6</code> | | zen6-coder | <code>zenlm/zen6-coder</code> @ <code>3f803c67</code> | 97,820,026,179 | other: qwen-community-1.0 | <code>zen6-coder</code> | | zen6-flash | <code>zenlm/zen6-flash</code> @ <code>b5e7a43c</code> | 16,769,698,285 | apache-2.0 | none |</p>
|
||
<p>GitHub has <code>zenlm/zen6</code>, <code>zenlm/zen6-coder</code> and <code>zenlm/zen6-flash</code>. They hold the card, plus a <code>training/</code> recipe for zen6 and zen6-coder, and no weights.</p>
|
||
<h3>§2 zen6</h3>
|
||
<p>The target model is a <code>Qwen3_5ForConditionalGeneration</code> (<code>model_type: qwen3_5</code>):</p>
|
||
<p>| Field | Value (<code>config.json</code>) | |:--|:--| | layers | 64: 48 <code>linear_attention</code> (Gated DeltaNet) and 16 <code>full_attention</code>, one full every 4 | | hidden, intermediate | 5,120 and 17,408 | | full attention | 24 query heads, 4 key/value heads, head dim 256, output gate | | linear attention | 16 key heads and 48 value heads, head dim 128, conv kernel 4 | | vocabulary | 248,320 | | MTP | 1 layer | | vision | 27-layer vision tower, hidden 1,152; image and video tokens; <code>language_model_only: false</code> | | positions | 262,144 native; <code>rope_type: yarn</code>, <code>factor: 4.0</code> → 1,048,576; <code>rope_theta</code> 10,000,000; <code>partial_rotary_factor</code> 0.25; interleaved mRoPE |</p>
|
||
<p>Quantization is NVIDIA ModelOpt 0.47.0.dev0 <code>MIXED_PRECISION</code>:</p>
|
||
<ul><li><strong>NVFP4</strong> (4-bit float weights and activations, group size 16) on 193 modules: every</li></ul>
|
||
<p> <code>gate_proj</code>, <code>up_proj</code> and <code>down_proj</code>, plus <code>lm_head</code>.</p>
|
||
<ul><li><strong>FP8</strong> (8-bit float weights and activations) on 208 modules: every attention projection,</li></ul>
|
||
<p> linear and full.</p>
|
||
<ul><li><strong>KV cache:</strong> FP8.</li><li><strong>Unquantized:</strong> MTP.</li></ul>
|
||
<p>The weights are three safetensors shards holding 2,194 tensors and 21,921,428,072 bytes: 1,842 language-model tensors, 333 vision, 15 MTP and 4 <code>lm_head</code>.</p>
|
||
<p>The drafter is <code>dflash2/</code>: <code>DFlash2DraftModel</code> in bfloat16, 3,848,817,896 bytes.</p>
|
||
<p>| Field | Value (<code>dflash2/config.json</code>) | |:--|:--| | layers | 5, <code>sliding_attention</code>, window 2,048 | | hidden | 5,120; 32 query heads, 8 key/value heads | | <code>block_size</code> | 8 | | <code>target_layer_ids</code> | 5, 19, 33, 47, 61 | | positions | 262,144, <code>rope_type: default</code> |</p>
|
||
<h3>§3 zen6-coder</h3>
|
||
<p>The zen6-coder GGUF headers give:</p>
|
||
<p>| Field | Value (GGUF header) | |:--|:--| | <code>general.architecture</code> | <code>qwen4exp</code>; <code>general.name</code> "Qwen3.8 Flash Next" | | blocks | 48, one full-attention block every 4 | | embedding | 2,560 | | attention | 24 heads, 2 key/value heads, key and value length 256; an indexer of 4 heads, key length 128, top-k 2,048 | | experts | 512, 10 used, expert FFN 640, shared-expert FFN 640 | | linear blocks | SSM conv kernel 4, state 128, 16 groups, inner 6,144 | | hyper-connections | 4, low rank 320 | | per-layer n-gram embedding | n-gram size 3; 16 hashed heads of about 20,000,000 entries each; <code>embedding_length_per_layer_input</code> 160 | | context | 262,144 | | quantization | Unsloth imatrix, file type IQ4_XS, 1,224 tensors across 3 splits |</p>
|
||
<p>The repository holds five files:</p>
|
||
<ul><li><code>UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-0000{1,2,3}-of-00003.gguf</code>, 93,682,584,224 bytes in</li></ul>
|
||
<p> total.</p>
|
||
<ul><li><code>MTP/mtp-Qwen3.8-Flash-Next-Q8_0.gguf</code>, 4,137,429,120 bytes. It is the same architecture with</li></ul>
|
||
<p> 49 blocks and <code>nextn_predict_layers: 1</code>.</p>
|
||
<ul><li><code>MTP/README.md</code>.</li></ul>
|
||
<p>The card states 180B total parameters, 125B in the base model and 6B active per token. The headers are consistent with that but do not prove it, and this HIP does not assert it.</p>
|
||
<h3>§4 zen6-flash</h3>
|
||
<p>The PQ2_0 GGUF header gives:</p>
|
||
<p>| Field | Value (PQ2_0 GGUF header) | |:--|:--| | <code>general.architecture</code> | <code>qwen35</code> | | blocks | 64, one full-attention block every 4 | | embedding, FFN | 5,120 and 17,408 | | attention | 24 heads, 4 key/value heads, key and value length 256 | | context | 262,144 | | tensors | 851: 402 of GGML type 142, 96 of BF16, 353 of F32 | | transform | a Walsh–Hadamard pre-rotation over 401 weights, block 1,024, recorded under <code>prism.hadamard.*</code> |</p>
|
||
<p>A loader MUST implement GGML type 142 and apply the recorded Hadamard transform before these weights mean anything.</p>
|
||
<p>| File | Bytes | |:--|--:| | <code>Ternary-Bonsai-2-27B-PQ2_0.gguf</code> | 7,206,168,928 | | <code>Ternary-Bonsai-2-27B-PTQ1_0.gguf</code> | 5,946,648,928 | | <code>Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf</code> | 629,246,976 | | <code>Ternary-Bonsai-2-27B-mmproj-BF16.gguf</code> | 931,145,856 | | <code>Bonsai-2-27B-DFlash2-Q8_0.gguf</code> | 2,056,415,104 | | <code>SHA256SUMS</code> | 65,916 |</p>
|
||
<h3>§5 Provenance</h3>
|
||
<p>Files match their upstream when the LFS SHA-256 or git blob id is equal. The comparison was run against each upstream's file listing from the same API.</p>
|
||
<p>| Checkpoint | Upstream | What matches | What differs | |:--|:--|:--|:--| | zen6 | <code>RadixArk/Qwen3.8-27B-NVFP4</code> (base <code>Qwen/Qwen3.8-27B</code> @ <code>e13a4f0e</code>) | 19 files: all shards, the index, the tokenizer files, the chat template, and the conversion, audit and qualification receipts | <code>config.json</code>, in three fields (below); the README; no <code>LICENSE</code>, which the upstream ships | | zen6 | <code>incoai/Qwen3.8-27B-DFlash2</code> | <code>config.json</code> and <code>model.safetensors</code>, placed under <code>dflash2/</code> | none | | zen6-coder | <code>unsloth/Qwen3.8-Flash-Next-GGUF</code> (base <code>Qwen/Qwen3.8-Flash-Next</code>) | all five files | the README | | zen6-flash | <code>prism-ml/Ternary-Bonsai-2-27B-gguf</code> (base <code>Qwen/Qwen3.8-27B</code>) | the four model and projector files | the README | | zen6-flash | <code>ProCreations/Ternary-Bonsai-2-27B-DFlash2</code> | the drafter and <code>SHA256SUMS</code> | none |</p>
|
||
<p>The three fields that differ in zen6's <code>config.json</code>, all under <code>text_config.rope_parameters</code>:</p>
|
||
<ul><li><code>rope_type</code>: <code>default</code> → <code>yarn</code></li><li><code>factor</code>: absent → <code>4.0</code></li><li><code>original_max_position_embeddings</code>: absent → <code>262144</code></li></ul>
|
||
<p>All three checkpoints descend from Qwen3.8, so they meet HIP-0039 §3.3.</p>
|
||
<h3>§6 Measured</h3>
|
||
<p>The only receipt in any Zen6 repository is zen6's <code>qualification.json</code> with its <code>qualification-criteria.json</code>. Both are byte-identical to the upstream's.</p>
|
||
<p>| Field | Value | |:--|:--| | task | GSM8K, 1,283 correct of 1,319 = 0.9727, against a gate of 0.965 | | failures | 0 empty generations, 0 truncations, 0 request errors | | harness | <code>sgl-eval</code> @ <code>6690895</code> | | server | SGLang @ <code>2948168</code> | | hardware | 4 × NVIDIA GB300, tensor parallel 4 | | KV cache | FP8 E4M3; scales not calibrated, so they default to 1.0 | | speculation | NEXTN on the MTP head, resolved as EAGLE; final acceptance length 2.775 | | sampling | thinking on, temperature 1.0, top-p 0.95, top-k 20 |</p>
|
||
<p><strong>What the receipt covers.</strong> It measures the upstream checkpoint with its upstream config. No receipt covers the YaRN edit, any context above 262,144, or the DFlash 2 drafter.</p>
|
||
<p><strong>Unmeasured claims.</strong> The cards' throughput tables, DFlash 2 speedups, retention percentages and benchmark comparisons carry no receipt. This HIP relies on none of them.</p>
|
||
<p><strong>Serving.</strong> The live SKUs <code>zen6</code> and <code>zen6-coder</code> report a 1,000,000-token window on <code>/v1/models</code>. Under HIP-0039 §2.3, that window is the SKU's contract and says nothing about these files.</p>
|
||
<h3>§7 Runtimes</h3>
|
||
<p>The upstream receipt names only one runtime for these weights: SGLang @ <code>2948168</code> running zen6 (§6).</p>
|
||
<p><code>hanzoai/engine</code> @ <code>d25f18e</code> contains:</p>
|
||
<ul><li>a <code>qwen4exp</code> model (<code>hanzo-engine/src/models/qwen4exp.rs</code> and <code>quantized_qwen4exp.rs</code>);</li><li>a <code>--dflash <dir></code> draft option;</li><li>an <code>--mtp-model</code> option.</li></ul>
|
||
<p>Whether it loads any of these three checkpoints has not been measured.</p>
|
||
<h3>§8 Defects</h3>
|
||
<p>As of 2026-09-25. A card fix is a new revision of that repository, and the repository revisions in §1 move with it.</p>
|
||
<ol><li><strong>Licence files.</strong> None of the three repositories carries <code>LICENSE</code> or <code>NOTICE</code>, which</li></ol>
|
||
<p> HIP-0039 §3.2 requires.</p>
|
||
<ol><li><strong>zen6 card against <code>config.json</code>.</strong> Under HIP-0039 §3.5 the config is the fact.</li></ol>
|
||
<p> | Field | Card says | <code>config.json</code> says | |:--|:--|:--| | heads | 32 query, 8 key/value | 24 query, 4 key/value | | vocabulary | 152,064 | 248,320 | | target attention | "Sliding-Window Attention" | full attention; sliding windows are the drafter's | | drafter block | 3–5 tokens | <code>block_size: 8</code> | | drafter positions | a 1M position table | 262,144 | | modality | <code>pipeline_tag: text-generation</code> | ships a vision tower |</p>
|
||
<ol><li><strong>zen6-coder.</strong> The card describes a "Halogen W4B" format, and no such file is in the</li></ol>
|
||
<p> repository. The GitHub description says "38B-Class"; the card says 180B.</p>
|
||
<ol><li><strong>zen6-flash.</strong> The card lists the BF16 projector as 1.2 GB. The file is 931,145,856 bytes.</li><li><strong>Commands that cannot run.</strong> The <code>hanzo-engine serve</code> commands on the zen6 and zen6-coder</li></ol>
|
||
<p> cards pass <code>--context-window</code>, <code>--kv-cache-quant</code> and <code>--mtp</code>. The engine at <code>d25f18e</code> defines none of them.</p>
|
||
<ol><li><strong>Not HIP-0039 §3.4.</strong> None of the cards says it is a re-host, and none names the files it</li></ol>
|
||
<p> copied.</p>
|
||
<h2>Rationale</h2>
|
||
<ul><li><strong>Why re-host.</strong> It puts a current checkpoint under the Zen name the day upstream publishes</li></ul>
|
||
<p> one, with no training cost. Its price is that Zen6 open weights are exactly as good as their upstreams, and the cards have to say so (HIP-0039 §3.4). Training a Zen-specific checkpoint would be a later generation with its own receipt.</p>
|
||
<ul><li><strong>Why the files are the specification.</strong> <code>config.json</code>, the GGUF headers and LFS hashes are</li></ul>
|
||
<p> what a runtime reads and what a reader can re-check. Card prose is neither.</p>
|
||
<ul><li><strong>Why the defects are listed here.</strong> A HIP that repeated the cards would carry their errors.</li></ul>
|
||
<p> One that silently corrected them would hide that the published cards are wrong. Listing the defects lets each be fixed and struck through.</p>
|
||
<h2>Security Considerations</h2>
|
||
<ul><li><strong>Pin revisions.</strong> A loader SHOULD pin the revision in §1, or its successor, and check LFS</li></ul>
|
||
<p> SHA-256 against the upstream (§5).</p>
|
||
<ul><li><strong>zen6-flash <code>SHA256SUMS</code>.</strong> It is the drafter repository's, and it lists 481 files. Of the</li></ul>
|
||
<p> files zen6-flash ships, it covers only the drafter. It does not cover the PQ2_0, PTQ1_0 or projector files, and it names a <code>LICENSE</code> and <code>NOTICE</code> that zen6-flash does not ship.</p>
|
||
<ul><li><strong>YaRN beyond the receipt.</strong> The zen6 YaRN edit takes the model outside its measured regime. A</li></ul>
|
||
<p> runtime that applies YaRN statically scales every position, including the short prompts the receipt measured. A deployment that relies on either long context or short-prompt accuracy MUST measure its own.</p>
|
||
<ul><li><strong>Uncalibrated KV cache.</strong> The FP8 KV cache ran with uncalibrated scales, and the qualification</li></ul>
|
||
<p> record says so. This affects accuracy, not isolation.</p>
|
||
<ul><li><strong>Custom tensor types.</strong> zen6-flash stores weights as GGML type 142 behind a Hadamard</li></ul>
|
||
<p> transform. Only a runtime that implements that type decodes them; any other loader MUST refuse the file rather than guess.</p>
|
||
<ul><li><strong>Licence.</strong> zen6-coder is licensed qwen-community-1.0, not Apache-2.0. Its terms bind anyone</li></ul>
|
||
<p> who redistributes or serves it.</p>
|
||
<ul><li><strong>Remote code.</strong> No Zen6 repository ships Python, and none carries the <code>custom_code</code> tag.</li></ul>
|
||
<h2>References</h2>
|
||
<ul><li>HIP-0039 Zen; HIP-0043 inference engine; HIP-0510 Enso router; HIP-0511 model families.</li><li>https://huggingface.co/zenlm/zen6 · https://huggingface.co/zenlm/zen6-coder ·</li></ul>
|
||
<p> https://huggingface.co/zenlm/zen6-flash</p>
|
||
<ul><li>Upstreams: <code>RadixArk/Qwen3.8-27B-NVFP4</code>, <code>incoai/Qwen3.8-27B-DFlash2</code>,</li></ul>
|
||
<p> <code>unsloth/Qwen3.8-Flash-Next-GGUF</code>, <code>prism-ml/Ternary-Bonsai-2-27B-gguf</code>, <code>ProCreations/Ternary-Bonsai-2-27B-DFlash2</code>.</p>
|
||
<ul><li>https://github.com/zenlm/zen6 · https://github.com/zenlm/zen6-coder ·</li></ul>
|
||
<p> https://github.com/zenlm/zen6-flash</p>
|
||
<h2>Copyright</h2>
|
||
<p>Copyright and related rights waived via <a href="https://creativecommons.org/publicdomain/zero/1.0/">CC0</a>.</p>
|
||
<footer>
|
||
<div class="fmap">
|
||
<div><b>Build</b>
|
||
<a href="https://hanzo.app">hanzo.app</a> app builder
|
||
<a href="https://hanzo.codes">hanzo.codes</a> code with an agent
|
||
<a href="https://hanzoskills.com">hanzoskills.com</a> what an agent reads</div>
|
||
<div><b>Run</b>
|
||
<a href="https://console.hanzo.ai">console.hanzo.ai</a> cloud console
|
||
<a href="https://api.hanzo.ai">api.hanzo.ai</a> the API</div>
|
||
<div><b>Account</b>
|
||
<a href="https://hanzo.id">hanzo.id</a> sign in
|
||
<a href="https://pay.hanzo.ai">pay.hanzo.ai</a> top up
|
||
<a href="https://billing.hanzo.ai">billing.hanzo.ai</a> invoices and usage</div>
|
||
<div><b>Company</b>
|
||
<a href="https://hanzo.ai">hanzo.ai</a> canonical
|
||
<a href="https://hanzo.ventures">hanzo.ventures</a> ventures</div>
|
||
</div>
|
||
<div class="fend"><a href="/hip-0904.md">This page as plain text</a> · <a href="/llms.txt">Index for agents</a></div>
|
||
</footer>
|
||
</div></body></html> |