Compare DeepSeek Flash 1017 reported scores (#10)

- Compare DeepSeek Flash 1017 reported scores (ff45b46e5092ad2757a5bf93943161e6d76e2219)

NOTICE

@@ -26,3 +26,11 @@ No compiled inference executable or model weights are included in this Space.
 Qwen measurements use alphaBell Qwen Open, a post-trained derivative.
 Its separate checkpoint release uses alphaBell's research and noncommercial license.
 Parent architecture configurations are distinct from captured checkpoint identity.
+
+DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017 benchmark score references:
+https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
+https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
+The DeepSeek model repositories are MIT licensed, Copyright (c) 2023 DeepSeek:
+https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/LICENSE
+Only score facts and configuration checksums are included in the release lens;
+model weights and source configuration files are not bundled here.

README.md

@@ -14,6 +14,8 @@ models:
   - Qwen/Qwen3.8-27B
   - alphabell/alphaBell-Qwen3.8-27B-Open-GGUF
   - alphabell/Flint-Matched-GGUF
+  - deepseek-ai/DeepSeek-V4.1-Flash-1017
+  - deepseek-ai/DeepSeek-V4.1-Flash
 tags:
   - interpretability
   - visualization
@@ -25,6 +27,8 @@ tags:
   - flint
   - nanbeige
   - architecture
+  - deepseek
+  - benchmark
   - educational
 ---
 
@@ -34,10 +38,14 @@ tags:
 
 Explore two different architectures with real Q8_0 tensor captures: Flint's sparse expert routing and alphaBell Qwen Open's recurrent memory. Open a block, head or expert, then step through the tokens that used it.
 
-[**Open the explorer**](https://huggingface.co/spaces/alphabell/Model-Observatory) · [Nanbeige and Qwen architecture](https://alphabell-model-observatory.static.hf.space/architecture/index.html) · [Qwen view](https://alphabell-model-observatory.static.hf.space/?model=qwen) · [Source](https://huggingface.co/spaces/alphabell/Model-Observatory/tree/main) · [alphaBell](https://alphabell.com)
+[**Open the explorer**](https://huggingface.co/spaces/alphabell/Model-Observatory) · [DeepSeek Flash 1017 comparison](https://alphabell-model-observatory.static.hf.space/deepseek/index.html) · [Nanbeige and Qwen architecture](https://alphabell-model-observatory.static.hf.space/architecture/index.html) · [Qwen view](https://alphabell-model-observatory.static.hf.space/?model=qwen) · [Source](https://huggingface.co/spaces/alphabell/Model-Observatory/tree/main) · [alphaBell](https://alphabell.com)
 
 ![Model Observatory showing Flint attention heads and selected experts for one captured token.](preview.png)
 
+## DeepSeek Flash 1017: reported scores and published configs
+
+[**Explore the paired-score comparison**](https://alphabell-model-observatory.static.hf.space/deepseek/index.html). The pinned DeepSeek-V4.1-Flash-1017 model card reports higher scores than the original DeepSeek-V4.1-Flash on all 11 paired benchmarks. The two pinned releases publish byte-identical `config.json` and `inference/config.json` files. Scores are publisher-reported and have not been reproduced by alphaBell. Matching configs do not establish identical weights or explain the reported gains. [Source data and revisions](deepseek/data.json).
+
 ## Nanbeige and Qwen architecture
 
 [**Explore the layer map**](https://alphabell-model-observatory.static.hf.space/architecture/index.html) to compare Nanbeige4.3-33B's 52 full-attention layers with the Qwen3.8-27B text decoder's 48 recurrent linear and 16 full-attention layers. This comparison uses published configurations; the measured captures in this Space are Flint and Qwen.

deepseek/data.json

@@ -0,0 +1,28 @@
+{
+  "title": "DeepSeek V4.1 Flash: 1017 compared with the original",
+  "source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/README.md",
+  "original_source": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/314113f8a7dadb388215396ba5fbee2c610a04d6/README.md",
+  "releases": {
+    "new": "DeepSeek-V4.1-Flash-1017",
+    "original": "DeepSeek-V4.1-Flash"
+  },
+  "config_sha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879",
+  "inference_config_sha256": "2e84f45cf1dac8c7fcbb200e96667d4b913275690668ed496f24c7747207a809",
+  "revisions": {
+    "new": "1bae0e1088993b9b82de08f59de5be5a34e42170",
+    "original": "314113f8a7dadb388215396ba5fbee2c610a04d6"
+  },
+  "benchmarks": [
+    {"id":"hle","name":"HLE without tools","original":36.8,"new":42.6,"note":"Text-only HLE subset: 39.1 original, 44.0 in 1017. The main chart uses the full HLE without-tools pair."},
+    {"id":"terminal","name":"Terminal-Bench 4.0","original":31.2,"new":42.3,"note":"The publisher reports this result using the Claude Code harness."},
+    {"id":"deepswe","name":"DeepSWE v1.1","original":74.2,"new":77.0,"note":"The publisher reports this result using the Mini-SWE-V2 harness."},
+    {"id":"program","name":"ProgramBench","original":20.3,"new":29.1},
+    {"id":"nl2repo","name":"NL2Repo-Bench","original":65.4,"new":68.2},
+    {"id":"automation","name":"Automation-Bench v1.0.6","original":54.8,"new":60.5},
+    {"id":"agents","name":"Agents’ Last Exam","original":31.8,"new":35.4},
+    {"id":"apex","name":"Apex Bench v1.1","original":64.6,"new":72.1},
+    {"id":"chartography","name":"Chartography with tools","original":78.9,"new":83.0,"with_tools":true},
+    {"id":"babyvision","name":"BabyVision with tools","original":89.6,"new":92.5,"with_tools":true},
+    {"id":"zerobench","name":"ZeroBench-main with tools","original":49.0,"new":51.0,"with_tools":true}
+  ]
+}

deepseek/index.html

@@ -0,0 +1,24 @@
+<!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="theme-color" content="#111c2a">
+  <meta name="description" content="An interactive, source-pinned comparison of the published DeepSeek V4.1 Flash 1017 and original benchmark scores and configuration files.">
+  <title>DeepSeek Flash 1017: same blueprint, new scores | alphaBell</title>
+  <link rel="icon" href="../favicon.svg">
+  <link rel="stylesheet" href="release.css">
+  <script src="release.js" defer></script>
+</head>
+<body>
+  <header class="topbar"><a class="brand" href="../index.html" aria-label="Model Observatory home"><span class="mark">α</span><span>alphaBell <span class="slash">/</span> observatory</span></a><a class="back" href="../index.html">Explore the Observatory <span aria-hidden="true">↗</span></a></header>
+  <main>
+    <section class="hero" aria-labelledby="title"><p class="eyebrow">Release lens <span>02 / source-pinned</span></p><h1 id="title">Same published blueprint.<br><em>New reported scores.</em></h1><p class="lead">DeepSeek V4.1 Flash 1017 and its predecessor publish byte-identical model and inference configurations. The 1017 model card reports higher scores on every one of its 11 paired benchmarks. Explore each pair below.</p><div class="scope"><span class="dot" aria-hidden="true"></span><p><strong>Reading the evidence</strong> Scores come from DeepSeek's model card; alphaBell has not rerun the benchmarks. Equal configuration files do not establish equal weights or explain the changes.</p></div></section>
+    <section class="headline" aria-label="Comparison summary"><div><strong>11<span>/11</span></strong><p>paired scores higher in 1017</p></div><div><strong id="largest-delta">+11.1</strong><p>largest reported score-point gain</p></div><div><strong>2<span>/2</span></strong><p>published configuration files identical</p></div></section>
+    <section class="compare" aria-labelledby="compare-title"><div class="section-head"><p class="eyebrow">01 / Reported benchmarks</p><h2 id="compare-title">What changed on the card?</h2><p>Compare the two releases within each row. Each benchmark has its own evaluation method, so scores across rows are not directly comparable.</p></div><div class="controls" role="group" aria-label="Benchmark controls"><div class="segmented" role="group" aria-label="Filter benchmarks"><button type="button" data-filter="all" aria-pressed="true">All 11</button><button type="button" data-filter="tools" aria-pressed="false">With tools</button><button type="button" data-filter="other" aria-pressed="false">Other</button></div><label for="sort">Order <select id="sort"><option value="card">Model card</option><option value="gain">Largest gain</option></select></label></div><div class="legend"><span><i class="swatch old"></i>Original</span><span><i class="swatch new"></i>1017</span><span class="axis-note">Scale: reported score, 0 to 100</span></div><div id="benchmarks" class="benchmark-list" aria-live="polite"></div><p class="benchmark-footnote">HLE's text-only subset is reported separately in the source: 39.1 for the original and 44.0 for 1017. Terminal-Bench 4.0 uses a Claude Code harness; DeepSWE v1.1 uses Mini-SWE-V2. All figures are publisher-reported.</p></section>
+    <section class="blueprint" aria-labelledby="blueprint-title"><div class="section-head"><p class="eyebrow">02 / Configuration check</p><h2 id="blueprint-title">What stayed in the published files?</h2><p>The pinned releases have the same byte-level SHA-256 digest for each of these two files. Expand a row to inspect both exact source links and the full digest.</p></div><details><summary><span>Model config</span><strong>Identical bytes</strong><span class="chevron" aria-hidden="true">+</span></summary><div class="detail"><code>8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879</code><div class="source-links"><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/config.json" target="_blank" rel="noopener noreferrer">1017 config ↗</a><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/314113f8a7dadb388215396ba5fbee2c610a04d6/config.json" target="_blank" rel="noopener noreferrer">Original config ↗</a></div></div></details><details><summary><span>Inference config</span><strong>Identical bytes</strong><span class="chevron" aria-hidden="true">+</span></summary><div class="detail"><code>2e84f45cf1dac8c7fcbb200e96667d4b913275690668ed496f24c7747207a809</code><div class="source-links"><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/inference/config.json" target="_blank" rel="noopener noreferrer">1017 inference config ↗</a><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/314113f8a7dadb388215396ba5fbee2c610a04d6/inference/config.json" target="_blank" rel="noopener noreferrer">Original inference config ↗</a></div></div></details><p class="caveat">A matching config describes the same published structure and settings. It says nothing about whether the weights, training data, post-training process or benchmark setup changed.</p></section>
+    <section class="sources" aria-labelledby="sources-title"><div class="section-head"><p class="eyebrow">Pinned sources</p><h2 id="sources-title">Check the primary evidence</h2></div><div class="source-cards"><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/1bae0e1088993b9b82de08f59de5be5a34e42170/README.md" target="_blank" rel="noopener noreferrer"><span>DeepSeek model card</span><strong>1017 release ↗</strong><small>Revision 1bae0e1</small></a><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/314113f8a7dadb388215396ba5fbee2c610a04d6/README.md" target="_blank" rel="noopener noreferrer"><span>DeepSeek model card</span><strong>Original release ↗</strong><small>Revision 314113f</small></a></div></section>
+  </main>
+  <footer><span>alphaBell / Model Observatory</span><a href="../architecture/index.html">Compare Nanbeige and Qwen architecture ↗</a></footer>
+</body>
+</html>

deepseek/release.css

@@ -0,0 +1 @@
+:root{font-family:Inter,ui-sans-serif,system-ui,-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif;color:#152739;background:#f5f8fa;font-synthesis:none}*{box-sizing:border-box}html{scroll-behavior:smooth}body{margin:0}button,select{font:inherit}a{color:inherit}a:focus-visible,button:focus-visible,select:focus-visible,summary:focus-visible{outline:3px solid #ffbe58;outline-offset:3px}.topbar{height:70px;display:flex;align-items:center;justify-content:space-between;padding:0 max(24px,calc((100vw - 1120px)/2));background:#111c2a;color:#fff}.brand{text-decoration:none;display:flex;align-items:center;gap:10px;font-size:14px;font-weight:650;letter-spacing:.01em}.mark{width:30px;height:30px;display:grid;place-items:center;background:#70e4c8;color:#0d2931;border-radius:9px;font-size:25px;line-height:1}.slash{color:#79aeb9;padding:0 2px}.back{color:#b9cbd1;text-decoration:none;font-size:13px}.back:hover,.source-links a:hover,footer a:hover{text-decoration:underline}main{max-width:1120px;margin:auto;padding:0 24px 104px}.hero{padding:72px 0 50px}.eyebrow{color:#1b7b74;text-transform:uppercase;letter-spacing:.15em;font-size:11px;font-weight:800;margin:0 0 17px}.eyebrow span{color:#8aa19d;margin-left:16px}.hero h1{font-size:clamp(44px,6.3vw,86px);line-height:1.045;letter-spacing:-.065em;margin:0;max-width:950px;font-weight:760}.hero h1 em{font-style:normal;color:#138774}.lead{font-size:clamp(17px,2vw,21px);line-height:1.52;max-width:770px;color:#465c6a;margin:27px 0 30px}.scope{display:flex;gap:15px;align-items:flex-start;max-width:820px;padding:17px 20px;background:#e6f4f0;border:1px solid #bce1d5;border-radius:13px}.scope p{margin:0;line-height:1.5;color:#35565a;font-size:14px}.scope strong{color:#104f47;margin-right:5px}.dot{width:9px;height:9px;border-radius:50%;background:#159879;margin-top:7px;flex:none;box-shadow:0 0 0 4px #c1e9dc}.headline{display:grid;grid-template-columns:repeat(3,1fr);gap:13px;margin:8px 0 88px}.headline>div{min-height:145px;padding:23px 26px;background:#111c2a;border-radius:16px;color:#fff}.headline strong{font-size:45px;letter-spacing:-.065em;line-height:1}.headline strong span{font-size:25px;color:#83b4bd}.headline p{color:#b5c8cf;font-size:14px;margin:13px 0 0}.section-head h2{font-size:clamp(31px,4vw,50px);letter-spacing:-.055em;line-height:1.1;margin:0 0 13px}.section-head>p:last-child{max-width:730px;color:#57707c;line-height:1.55;margin:0;font-size:16px}.compare,.blueprint,.sources{margin-top:87px}.controls{display:flex;align-items:center;justify-content:space-between;gap:14px;flex-wrap:wrap;margin:32px 0 24px}.segmented{display:inline-flex;gap:4px;padding:4px;background:#e4edf0;border-radius:11px}.segmented button{padding:9px 15px;border:0;border-radius:8px;background:transparent;color:#4b6470;font-weight:650;cursor:pointer}.segmented button[aria-pressed="true"]{background:#fff;color:#155f57;box-shadow:0 1px 4px #162d3222}.controls label{display:flex;gap:10px;align-items:center;color:#57707c;font-size:13px;font-weight:650}.controls select{border:1px solid #c6d4d9;border-radius:9px;background:#fff;color:#183242;padding:9px 30px 9px 11px}.legend{display:flex;gap:22px;align-items:center;color:#57707c;font-size:12px;margin-bottom:12px}.legend span{display:inline-flex;align-items:center;gap:7px}.swatch{width:10px;height:10px;border-radius:3px;display:inline-block}.swatch.old{background:#a7b9c5}.swatch.new{background:#19a88a}.legend .axis-note{margin-left:auto}.benchmark-list{display:grid;gap:10px}.benchmark{display:grid;grid-template-columns:minmax(168px,1fr) minmax(260px,1.55fr) 90px;align-items:center;gap:24px;background:#fff;border:1px solid #dce7ea;border-radius:13px;padding:18px 22px;box-shadow:0 3px 14px #24404b09}.benchmark h3{font-size:16px;margin:0 0 4px;line-height:1.3}.benchmark .note{font-size:12px;color:#657b85;line-height:1.45;margin:0}.barset{display:grid;gap:7px}.barrow{display:grid;grid-template-columns:31px 1fr 39px;gap:9px;align-items:center;color:#657b85;font-size:11px;font-weight:700}.track{height:9px;border-radius:8px;background:#ecf2f4;overflow:hidden}.bar{height:100%;display:block;border-radius:8px}.bar.old{background:#a7b9c5}.bar.new{background:#19a88a}.score{font-variant-numeric:tabular-nums;text-align:right;font-size:13px;color:#314e5a}.gain{justify-self:end;color:#0a8069;font-size:22px;font-weight:780;letter-spacing:-.04em;font-variant-numeric:tabular-nums;white-space:nowrap}.gain small{display:block;text-align:right;letter-spacing:0;font-weight:600;font-size:11px;color:#67827f}.benchmark-footnote{margin:21px 0 0;max-width:840px;color:#657b85;font-size:13px;line-height:1.6}.blueprint details{background:#fff;border:1px solid #dce7ea;border-radius:12px;margin-top:10px;overflow:hidden}.blueprint details:first-of-type{margin-top:30px}.blueprint summary{display:flex;align-items:center;gap:20px;cursor:pointer;padding:21px 22px;list-style:none;font-weight:720}.blueprint summary::-webkit-details-marker{display:none}.blueprint summary strong{margin-left:auto;color:#138774;font-size:13px}.chevron{color:#829aa2;font-size:24px;line-height:1}.blueprint details[open] .chevron{transform:rotate(45deg)}.detail{border-top:1px solid #e3ebed;padding:17px 22px 22px}.detail code{overflow-wrap:anywhere;font-size:12px;color:#4f6770}.source-links{display:flex;gap:18px;flex-wrap:wrap;margin-top:14px}.source-links a{font-size:13px;color:#086b5b;font-weight:700}.caveat{max-width:850px;line-height:1.55;color:#617984;font-size:14px;margin:20px 0 0}.source-cards{display:grid;grid-template-columns:1fr 1fr;gap:13px;margin-top:28px}.source-cards a{display:grid;gap:10px;text-decoration:none;background:#dff1ec;border:1px solid #b7dcd0;border-radius:14px;padding:24px;transition:transform .2s,box-shadow .2s}.source-cards a:hover{transform:translateY(-2px);box-shadow:0 8px 20px #17483a17}.source-cards span{font-size:12px;color:#48736a;text-transform:uppercase;letter-spacing:.09em;font-weight:700}.source-cards strong{font-size:19px;color:#075c50}.source-cards small{font-size:12px;color:#66857b}footer{display:flex;justify-content:space-between;gap:16px;padding:24px max(24px,calc((100vw - 1120px)/2));background:#111c2a;color:#a9c0c8;font-size:12px}footer a{text-decoration:none;color:#d4e9e6}@media(max-width:760px){.topbar{height:62px}.back{font-size:0}.back span{font-size:18px}.hero{padding:53px 0 35px}.headline{grid-template-columns:1fr;gap:8px;margin-bottom:65px}.headline>div{min-height:auto;padding:18px 22px;display:flex;align-items:center;gap:22px}.headline strong{font-size:39px;min-width:105px}.headline p{margin:0}.benchmark{grid-template-columns:1fr auto;gap:13px 15px}.barset{grid-column:1/-1;grid-row:2}.gain{grid-column:2;grid-row:1}.legend .axis-note{display:none}.source-cards{grid-template-columns:1fr}.compare,.blueprint,.sources{margin-top:65px}}@media(max-width:390px){main{padding:0 16px 72px}.topbar,footer{padding-left:16px;padding-right:16px}.brand{font-size:12px}.eyebrow{font-size:10px}.hero h1{font-size:42px}.controls{align-items:stretch}.segmented{width:100%}.segmented button{flex:1;padding:9px 8px;font-size:12px}.benchmark{padding:16px}.benchmark h3{font-size:14px}.gain{font-size:20px}.headline p{font-size:12px}footer{display:grid}}

deepseek/release.js

@@ -0,0 +1,34 @@
+'use strict';
+(async function () {
+  const list = document.querySelector('#benchmarks');
+  const buttons = [...document.querySelectorAll('[data-filter]')];
+  const sort = document.querySelector('#sort');
+  const escape = value => String(value).replace(/[&<>"']/g, ch => ({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[ch]));
+  let data;
+  let filter = 'all';
+  try {
+    const response = await fetch('data.json');
+    if (!response.ok) throw new Error('Could not load the source-pinned score data.');
+    data = await response.json();
+  } catch (error) {
+    list.textContent = error.message;
+    return;
+  }
+  function render() {
+    let rows = data.benchmarks.filter(row => filter === 'all' || Boolean(row.with_tools) === (filter === 'tools'));
+    if (sort.value === 'gain') rows = [...rows].sort((a,b) => (b.new-b.original)-(a.new-a.original));
+    list.innerHTML = rows.map(row => {
+      const difference = (row.new-row.original).toFixed(1);
+      const note = row.note ? `<p class="note">${escape(row.note)}</p>` : '';
+      return `<article class="benchmark" id="${escape(row.id)}"><div><h3>${escape(row.name)}</h3>${note}</div><div class="barset" aria-label="Original ${row.original.toFixed(1)}, 1017 ${row.new.toFixed(1)}"><div class="barrow"><span>Old</span><div class="track"><span class="bar old" style="width:${row.original}%"></span></div><span class="score">${row.original.toFixed(1)}</span></div><div class="barrow"><span>1017</span><div class="track"><span class="bar new" style="width:${row.new}%"></span></div><span class="score">${row.new.toFixed(1)}</span></div></div><div class="gain">+${difference}<small>score points</small></div></article>`;
+    }).join('');
+  }
+  buttons.forEach(button => button.addEventListener('click', () => {
+    filter = button.dataset.filter;
+    buttons.forEach(item => item.setAttribute('aria-pressed', String(item === button)));
+    render();
+  }));
+  sort.addEventListener('change', render);
+  document.querySelector('#largest-delta').textContent = '+' + Math.max(...data.benchmarks.map(row => row.new-row.original)).toFixed(1);
+  render();
+})();

index.html

@@ -3,7 +3,7 @@
 <head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Flint &amp; Qwen Model Observatory | alphaBell</title><meta name="description" content="Explore Flint expert routing and Qwen recurrent memory with real Q8 tensor captures. Replay every token, inspect every block and import your own trace."><meta property="og:title" content="Model Observatory by alphaBell"><meta property="og:description" content="Two architectures. Real tensor captures. Every token."><meta property="og:image" content="https://huggingface.co/spaces/alphabell/Model-Observatory/resolve/main/cover.png"><meta name="theme-color" content="#176b59"><link rel="icon" href="favicon.svg"><link rel="stylesheet" href="style.css"></head>
 <body>
 <header class="models-header"><div class="header-start"><a class="brand" href="./"><span class="brand-icon">α</span><span>alphaBell<span class="brand-divider">/</span><b>observatory</b></span></a><label class="model-control"><span>MODEL</span><select id="modelSelect" aria-label="Model" disabled><option>Recursive Flint</option></select></label></div><div class="header-end"><span class="engine-status" id="engineStatus"><i></i>Loading capture</span><button class="icon-button" id="aboutButton" title="How to read the model" aria-label="How to read the model">?</button></div></header>
-<section class="overview"><div><div class="eyebrow" id="modelEyebrow">MODEL OBSERVATORY</div><h1>Inside Flint &amp; Qwen.</h1><a class="experiment-entry" href="experiments/index.html">Try the five experiments ↗</a><a class="experiment-entry architecture-entry" href="architecture/index.html">Compare Nanbeige and Qwen architecture ↗</a></div><div class="stats" id="modelStats"></div></section>
+<section class="overview"><div><div class="eyebrow" id="modelEyebrow">MODEL OBSERVATORY</div><h1>Inside Flint &amp; Qwen.</h1><a class="experiment-entry" href="experiments/index.html">Try the five experiments ↗</a><a class="experiment-entry architecture-entry" href="architecture/index.html">Compare Nanbeige and Qwen architecture ↗</a><a class="experiment-entry architecture-entry" href="deepseek/index.html">Explore DeepSeek Flash 1017 scores ↗</a></div><div class="stats" id="modelStats"></div></section>
 <main>
 <aside class="input-panel">
  <section class="prompt-section"><div class="section-heading"><h2>01 <span>Measured capture</span></h2><span class="small-tag" id="quantization">Q8_0</span></div><p id="traceOrigin" class="trace-origin">Opening example capture…</p><label class="sr-only" for="prompt" id="promptLabel">Captured prompt</label><textarea id="prompt" readonly spellcheck="false">Explain why the sky is blue in two sentences.</textarea><p class="capture-settings" id="captureSettings"></p><button id="runButton" class="primary" disabled>Loading capture…</button><p id="runNote" class="run-note" role="status">Structure is ready while the capture loads.</p><details class="capture-help"><summary>Use your own capture</summary><p>Import a compatible Flint or Qwen trace JSON. It stays in your browser. This public viewer replays measured captures; new prompts need a separate capture engine.</p><a href="https://huggingface.co/spaces/alphabell/Model-Observatory/blob/main/TRACE-FORMAT.md" target="_blank" rel="noreferrer">Trace format ↗</a></details></section>

sources.json

@@ -12,6 +12,16 @@
       "sha": "c1093c501cc6460427c0ac634db7223d0450c0ee",
       "config": "https://huggingface.co/Nanbeige/Nanbeige4.3-33B/resolve/c1093c501cc6460427c0ac634db7223d0450c0ee/config.json",
       "bundled_config_sha256": "ee6db71f6de1bc2d78c0adc543e339c44e9340c1e3f474285f3144d2d3cd3df9"
+    },
+    "deepseek-ai/DeepSeek-V4.1-Flash-1017": {
+      "sha": "1bae0e1088993b9b82de08f59de5be5a34e42170",
+      "config_sha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879",
+      "inference_config_sha256": "2e84f45cf1dac8c7fcbb200e96667d4b913275690668ed496f24c7747207a809"
+    },
+    "deepseek-ai/DeepSeek-V4.1-Flash": {
+      "sha": "314113f8a7dadb388215396ba5fbee2c610a04d6",
+      "config_sha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879",
+      "inference_config_sha256": "2e84f45cf1dac8c7fcbb200e96667d4b913275690668ed496f24c7747207a809"
     }
   },
   "capture_engine": {

tests/deepseek-release.test.cjs

@@ -0,0 +1,47 @@
+'use strict';
+const test = require('node:test');
+const assert = require('node:assert/strict');
+const fs = require('node:fs');
+const path = require('node:path');
+const root = path.resolve(__dirname, '..');
+const data = JSON.parse(fs.readFileSync(path.join(root, 'deepseek/data.json'), 'utf8'));
+const sources = JSON.parse(fs.readFileSync(path.join(root, 'sources.json'), 'utf8'));
+
+test('release lens preserves every paired publisher score and important footnotes', () => {
+  const pairs = [
+    ['hle',36.8,42.6], ['terminal',31.2,42.3], ['deepswe',74.2,77.0],
+    ['program',20.3,29.1], ['nl2repo',65.4,68.2], ['automation',54.8,60.5],
+    ['agents',31.8,35.4], ['apex',64.6,72.1], ['chartography',78.9,83.0],
+    ['babyvision',89.6,92.5], ['zerobench',49.0,51.0]
+  ];
+  assert.deepEqual(data.benchmarks.map(row => [row.id,row.original,row.new]), pairs);
+  assert.ok(data.benchmarks.every(row => row.new > row.original));
+  assert.equal(Math.max(...data.benchmarks.map(row => row.new - row.original)).toFixed(1), '11.1');
+  assert.equal(data.benchmarks.filter(row => row.with_tools).length, 3);
+  assert.match(data.benchmarks[0].note, /39\.1 original, 44\.0 in 1017/);
+  assert.match(data.benchmarks[1].note, /Claude Code/);
+  assert.match(data.benchmarks[2].note, /Mini-SWE-V2/);
+});
+
+test('revisions and config digests agree with the public source manifest', () => {
+  const current = sources.configurations['deepseek-ai/DeepSeek-V4.1-Flash-1017'];
+  const prior = sources.configurations['deepseek-ai/DeepSeek-V4.1-Flash'];
+  assert.equal(current.sha, data.revisions.new);
+  assert.equal(prior.sha, data.revisions.original);
+  for (const config of [current,prior]) {
+    assert.equal(config.config_sha256,data.config_sha256);
+    assert.equal(config.inference_config_sha256,data.inference_config_sha256);
+  }
+  assert.match(data.source, new RegExp(data.revisions.new));
+  assert.match(data.original_source, new RegExp(data.revisions.original));
+});
+
+test('release page exposes primary sources, scope, and included assets', () => {
+  const html = fs.readFileSync(path.join(root,'deepseek/index.html'),'utf8');
+  const readme = fs.readFileSync(path.join(root,'README.md'),'utf8');
+  assert.match(html,/publisher-reported/i);
+  assert.match(html,/equal configuration files do not establish equal weights/i);
+  assert.match(html,/DeepSeek-V4\.1-Flash-1017\/blob\/1bae0e1088993b9b82de08f59de5be5a34e42170\/README\.md/);
+  assert.match(readme,/deepseek\/index\.html/);
+  for (const asset of ['release.css','release.js','data.json']) assert.ok(fs.statSync(path.join(root,'deepseek',asset)).isFile());
+});