<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>3B vs 671B: Nearly Identical AIME26 Scores, Different Tests (Preprint)</title>
        <link>https://peertube.watch/videos/watch/d323f67c-b4d2-4464-a3de-0b4ceb690731</link>
        <description>Peace Call. Mr. Peace is an AI-made character; voice is AI-generated; script written with AI assistance. No affiliate links in this video. How this site is run: https://getpeacecall.com/verify/?utm_source=peertube&amp;utm_medium=video&amp;utm_campaign=underdog-vibethinker According to a vendor-authored preprint, VibeThinker-3B scored 94.3 vs DeepSeek V3.2's 94.2 on AIME26 (2026 math contest problems, mean pass@1) - a 0.1 gap, which is not a win. One close math score doesn't prove equal overall ability. It clearly won IFBench (instruction following), 74.5 vs 60.7. But across the 10 benchmarks in the report's Table 2, DeepSeek V3.2 (671B, about 224x larger) scored higher on 7, including GPQA-Diamond science knowledge, 82.4 vs 70.2. Source: Xu et al. (Sina Weibo), 'VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models', arXiv:2606.16140v1 - PREPRINT technical report (Jun 15, 2026), not yet peer reviewed. https://arxiv.org/abs/2606.16140 . The authors built the small model; the comparison model's scores were collected from public reports, not rerun side by side. Numbers from Table 2 (without test-time scaling). Model names are trademarks of their owners; not affiliated with Weibo, DeepSeek or the authors. Charts drawn by code from the paper's numbers. Mr. Peace is an AI-made character with an AI voice. Paper: https://arxiv.org/abs/2606.16140 (version v1, posted Jun 15, 2026). AI Disclosure: Mr. Peace is a fictional AI character using a synthetic voice. This video is AI-assisted and does not depict a real person speaking. Research Disclosure: Based on a research preprint that has not been peer-reviewed. Results apply to the reported benchmarks and experimental conditions, not overall model superiority. Trademark Disclosure: Product, model, and game names are used for identification and commentary. Trademarks belong to their respective owners. Peace Call is not affiliated with or endorsed by those companies.</description>
        <lastBuildDate>Tue, 03 Nov 2026 15:54:08 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://peertube.watch</generator>
        <image>
            <title>3B vs 671B: Nearly Identical AIME26 Scores, Different Tests (Preprint)</title>
            <url>https://peertube.watch/lazy-static/avatars/3b284627-b75f-409e-ac2e-10f4c2946f43.png</url>
            <link>https://peertube.watch/videos/watch/d323f67c-b4d2-4464-a3de-0b4ceb690731</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://peertube.watch/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://peertube.watch/feeds/video-comments.xml?videoId=d323f67c-b4d2-4464-a3de-0b4ceb690731" rel="self" type="application/rss+xml"/>
    </channel>
</rss>