<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>DeviceMark — On-Device LLM Leaderboard changelog</title>
  <link>https://devicemark.github.io/changelog.html</link>
  <atom:link href="https://devicemark.github.io/feed.xml" rel="self" type="application/rss+xml"/>
  <description>Board updates: new models, new device measurements, methodology changes.</description>
  <language>en</language>

  <item>
    <title>Scoring made bit-reproducible; two upstream IFEval items disclosed</title>
    <link>https://devicemark.github.io/methodology.html#ci</link>
    <guid isPermaLink="false">devicemark-2026-07-14-determinism</guid>
    <pubDate>Tue, 14 Jul 2026 02:00:00 GMT</pubDate>
    <description>The official IFEval checkers are non-reproducible unseeded (langdetect, hash order, and upstream items 1122/1129 whose non-alphabet letters the checker replaces with a random letter). All pinned; the board now regenerates bit-identically from the published raw outputs. Values shifted at most 0.3pp, inside every CI.</description>
  </item>

  <item>
    <title>v0 launch — 8 entries, full battery, device-measured</title>
    <link>https://devicemark.github.io/</link>
    <guid isPermaLink="false">devicemark-2026-07-13-v0</guid>
    <pubDate>Sun, 13 Jul 2026 06:00:00 GMT</pubDate>
    <description>8 entries (5 open-model vendors + Apple's built-in Foundation Model), full 596-item battery, 95% CIs, 7/8 rows device-measured on iPhone 17 Pro. Findings: five-way statistical tie at the top; Google's official QAT int4 at parity on MMLU-Pro/MATH; quantization loss not monotone with size. Dataset: https://huggingface.co/datasets/devicemark/results</description>
  </item>

  <item>
    <title>Per-model badges</title>
    <link>https://devicemark.github.io/changelog.html</link>
    <guid isPermaLink="false">devicemark-2026-07-13-badges</guid>
    <pubDate>Sun, 13 Jul 2026 09:00:00 GMT</pubDate>
    <description>Embeddable per-row badges at /badge/&lt;slug&gt;.svg — intelligence composite + device tok/s, regenerated on every deploy.</description>
  </item>

</channel>
</rss>
