Files
zeekay c38d21b533 Point Benchmarks and Philosophy CTAs to platform.hanzo.ai
platform.hanzo.ai is the platform console (with cloud.hanzo.ai
redirecting to it at the edge). Point the remaining cloud.hanzo.ai
links on /benchmarks and /philosophy directly at https://platform.hanzo.ai.

Claude-Session: https://claude.ai/code/session_015XYKLix9PSyn6Jw2HWKprk
2026-11-09 12:16:29 -08:00

1 line
24 KiB
HTML
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
<div class="is_View _col-_platformweb_text-primar121 _fd-column _dsp-block _minH-100vh _bg-background"><main class="is_View _col-_platformweb_text-primar121 _fd-column _maxW-1080px _width-10037 _mr-auto _ml-auto _pt-64px _pr-24px _pl-24px _pb-96px _bg-background"><p class="is_Text _col-muted-foreg108704142 _dsp-block _mb-4px _fs-f-size-1 _lh-f-lineHeigh112922" data-slot="text"><a href="/benchmarks">Benchmarks</a> / <!-- -->Engine prefill on CUDA vs llama.cpp</p><span class="is_Text _col-muted-foreg108704142 _dsp-block _fs-f-size-1 _lh-f-lineHeigh112922 _ls-0--88px _tt-uppercase" data-slot="text">measured</span><h1 class="is_Text font_heading _fs-_sm_f-size-11 _fs-_lg_text-6xl _textWrap-_platformweb_balance _col-foreground _dsp-revert-laye114 _ws-inherit _ff-f-family _fs-f-size-10 _fw-497 _lh-1--19 _ls--0--012em _textWrap-balance _mt-8px _mb-16px" data-slot="display">Engine prefill on CUDA vs llama.cpp</h1><p class="is_Text _col-muted-foreg108704142 _dsp-revert-laye114 _ws-inherit _fs-f-size-3 _lh-1--65 _ls-0px _textWrap-pretty _maxW-720px" data-slot="lede">hanzo-engine against llama.cpp on CUDA, built against the same weights and run on the same box: prefill throughput at the prompt lengths where the whole 95% interval of the ratio lies above parity.</p><dl class="is_Text _col-text-primar121 _dsp-grid _gridTemplateColumns-minmax120px329255122 _rowGap-8px _columnGap-20px _mt-20px _mr-0px _mb-0px _ml-0px _maxW-760px _fs-f-size-2 _lh-f-lineHeigh112922" data-slot="text"><dt class="is_Text _col-muted-foreg108704142" data-slot="text">harness</dt><dd class="is_Text _col-inherit _mt-0px _mr-0px _mb-0px _ml-0px" data-slot="text"><a href="https://github.com/hanzoai/engine/tree/bench/dossier/bench-runs">hanzoai/engine · bench/dossier</a></dd><dt class="is_Text _col-muted-foreg108704142" data-slot="text">our paper</dt><dd class="is_Text _col-inherit _mt-0px _mr-0px _mb-0px _ml-0px" data-slot="text"><span class="is_Text _col-inherit" data-slot="text"><a href="https://papers.hanzo.ai/refutation-driven-perf">Refutation-Driven Performance Engineering</a></span></dd></dl><code class="is_Text font_mono _col-text-primar121 _dsp-block _mt-10px _pt-10px _pb-10px _pr-14px _pl-14px _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _btlr-6px _btrr-6px _bbrr-6px _bblr-6px _ox-auto _bg-card _ff-f-family _fs-f-size-1 _bbs-solid _bts-solid _bls-solid _brs-solid" data-slot="text">scripts/dossier.sh --backend cuda --model &lt;same.gguf&gt; -p 512,500,2048,4096 -n 128 -r 7</code><p class="is_Text font_mono _col-muted-foreg108704142 _dsp-block _mb-10px _ff-f-family _fs-f-size-1 _lh-f-lineHeigh112922 _mt-8px" data-slot="text">prints the table on this page — transcribed from the committed run files on branch bench/dossier, the pass of 2026-07-21</p><h2 class="is_Text font_heading _col-foreground _dsp-revert-laye114 _ws-inherit _ff-f-family _fs-f-size-7 _fw-497 _lh-1--2 _ls--0--01em _textWrap-balance _mt-56px _mb-8px" data-slot="title">What is measured</h2><p class="is_Text _col-text-second96872 _dsp-block _maxW-720px _lh-1--3563rem" data-slot="text">hanzo-engine and llama.cpp, built against the same weights, run on the same box, alternating. Prefill is tokens per second at a given prompt length, and the ratio is ours over theirs, so above 1.000 is a win. The models are Zen-1.7B at Q4_K_M and Zen-4B at Q4_K_M; 7 repetitions per cell, as its run records, with the first discarded as warmup, a Student-t 95% interval on each engine, and the ratio's interval formed by adding the two relative intervals in quadrature.</p><p class="is_Text _col-text-second96872 _dsp-block _maxW-720px _lh-1--3563rem" data-slot="text">The verdict is the harness's, not a reading of ours: <strong>WIN</strong> where the whole interval lies above 1.000 and neither engine's coefficient of variation exceeded 5%, which is how the harness says the box was quiet. Every cell below is one.</p><h2 class="is_Text font_heading _col-foreground _dsp-revert-laye114 _ws-inherit _ff-f-family _fs-f-size-7 _fw-497 _lh-1--2 _ls--0--01em _textWrap-balance _mt-56px _mb-8px" data-slot="title">Prefill on CUDA, cell by cell</h2><p class="is_Text font_mono _col-muted-foreg108704142 _dsp-block _mb-10px _ff-f-family _fs-f-size-1 _lh-f-lineHeigh112922" data-slot="text">3<!-- --> cells from <!-- -->2<!-- --> runs on branch <code>bench/dossier</code> · engine <!-- -->1.7.87<!-- --> · each run pins its engine release, its llama.cpp build and its model digest, and is linked below</p><div class="is_View _ox-auto _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _btlr-6px _btrr-6px _bbrr-6px _bblr-6px _bg-card _bbs-solid _bts-solid _bls-solid _brs-solid"><table class="is_Text _dsp-_platformweb_table _borderCollapse-_platformweb_collapse _fontVariantNumeric-_platformweb_tabular-num115 _col-text-primar121 _width-10037 _fs-f-size-2 _lh-f-lineHeigh112922" data-slot="text"><thead><tr><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-left _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">backend</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-left _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">model</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-left _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">quant</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-right _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">prompt tokens</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-right _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">llama.cpp t/s</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-right _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">hanzo t/s</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-right _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">ratio</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-right _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">95% interval</th><th class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-muted-foreg108704142 _text-left _pt-8px _pb-8px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _fw-500 _ws-nowrap _bbs-solid" data-slot="text">verdict</th></tr></thead><tbody><tr class="is_Text _dsp-_platformweb_table-row _col-inherit" data-slot="text"><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">CUDA</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Zen-1.7B</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Q4_K_M</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">500</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">7250.2</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">7961.3</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.098</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.053 – 1.143</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">WIN</td></tr><tr class="is_Text _dsp-_platformweb_table-row _col-inherit" data-slot="text"><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">CUDA</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Zen-1.7B</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Q4_K_M</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">4096</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">6871.3</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">7474.8</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.088</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.051 – 1.125</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">WIN</td></tr><tr class="is_Text _dsp-_platformweb_table-row _col-inherit" data-slot="text"><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">CUDA</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Zen-4B</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">Q4_K_M</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">2048</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">3592.5</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">3774.4</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.051</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-right _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">1.017 – 1.085</td><td class="is_Text _dsp-_platformweb_table-cell _fontVariantNumeric-_platformweb_tabular-num115 _col-inherit _text-left _pt-7px _pb-7px _pr-12px _pl-12px _borderBottomWidth-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _ws-nowrap _bbs-solid" data-slot="text">WIN</td></tr></tbody></table></div><h2 class="is_Text font_heading _col-foreground _dsp-revert-laye114 _ws-inherit _ff-f-family _fs-f-size-7 _fw-497 _lh-1--2 _ls--0--01em _textWrap-balance _mt-56px _mb-8px" data-slot="title">The work that produced it</h2><p class="is_Text _col-text-second96872 _dsp-block _maxW-720px _lh-1--3563rem" data-slot="text"><a href="https://papers.hanzo.ai/refutation-driven-perf">Refutation-Driven Performance Engineering</a> is the campaign that produced these kernels, written as a log of numbered hypotheses, each settled by a cheap decisive experiment. A long prefill is compute-bound, which is where that work landed.</p><div class="is_View _fd-column _mt-48px _pt-16px"><div class="is_View _text-_platformweb_center _fd-column _mt-56px _mb-24px _pt-36px _pb-36px _pr-32px _pl-32px _btlr-12px _btrr-12px _bbrr-12px _bblr-12px _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-border-stro3513 _brc-border-stro3513 _borderBottomColor-border-stro3513 _borderLeftColor-border-stro3513 _backgroundImage-linear-grad1889052070 _items-center _gap-16px _bbs-solid _bts-solid _bls-solid _brs-solid"><p class="is_Text font_mono _col-muted-foreg108704142 _dsp-block _mb-0px _ff-f-family _fs-f-size-1 _lh-f-lineHeigh112922 _mt-0px _mr-0px _ml-0px _ls-1--32px _tt-uppercase" data-slot="text">Hanzo AI Cloud</p><h3 class="is_Text _fs-_md_f-size-10 _lh-_md_2--5rem _col-text-primar121 _dsp-block _mt-0px _mr-0px _mb-0px _ml-0px _fs-f-size-7 _lh-1--6406rem _fw-600 _ls-f-letterSpa1360334199" data-slot="text">Build on Hanzo Cloud</h3><div class="is_View _fd-row _fwr-wrap _gap-12px _mt-8px _justify-center"><a class="is_Text _col-background _dsp-inline-flex _items-center _gap-8px _pt-12px _pb-12px _pr-24px _pl-24px _btlr-9999px _btrr-9999px _bbrr-9999px _bblr-9999px _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-transparent _brc-transparent _borderBottomColor-transparent _borderLeftColor-transparent _bg-text-primar121 _fs-f-size-3 _lh-f-lineHeigh112922 _fw-600 _bbs-solid _bts-solid _bls-solid _brs-solid" data-slot="text" href="https://platform.hanzo.ai">Start on Hanzo Cloud<svg aria-hidden="true" data-i18n-mirror="true" fill="none" height="14" stroke="var(--background)" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" style="color:var(--background);flex-shrink:0" viewbox="0 0 24 24" width="14"><path d="M5 12h14" stroke="var(--background)"></path><path d="m12 5 7 7-7 7" stroke="var(--background)"></path></svg></a><a class="is_Text _col-text-primar121 _dsp-inline-flex _items-center _gap-8px _pt-12px _pb-12px _pr-24px _pl-24px _btlr-9999px _btrr-9999px _bbrr-9999px _bblr-9999px _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-border-stro3513 _brc-border-stro3513 _borderBottomColor-border-stro3513 _borderLeftColor-border-stro3513 _bg-transparent _fs-f-size-3 _lh-f-lineHeigh112922 _fw-500 _bbs-solid _bts-solid _bls-solid _brs-solid" data-slot="text" href="/pricing">View Models &amp; Pricing</a></div></div><div class="is_View _fd-column _btw-1px _btc-edge _brc-edge _borderBottomColor-edge _borderLeftColor-edge _pt-32px _bts-solid"><div class="is_View _fd-row _fwr-wrap _justify-space-betwe3241 _items-center _gap-16px _mb-32px _pt-16px _pb-16px _pr-20px _pl-20px _bg-card _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-border _brc-border _borderBottomColor-border _borderLeftColor-border _btlr-8px _btrr-8px _bbrr-8px _bblr-8px _bbs-solid _bts-solid _bls-solid _brs-solid"><a class="is_View _fd-column _gap-2px _items-flex-start" href="/benchmarks/sandbox"><div class="is_View _fd-row _items-center _gap-4px"><svg aria-hidden="true" data-i18n-mirror="true" fill="none" height="12" stroke="var(--muted-foreground)" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" style="color:var(--muted-foreground);flex-shrink:0" viewbox="0 0 24 24" width="12"><path d="m12 19-7-7 7-7" stroke="var(--muted-foreground)"></path><path d="M19 12H5" stroke="var(--muted-foreground)"></path></svg><span class="is_Text _col-muted-foreg108704142 _fs-f-size-1 _lh-f-lineHeigh112922 _tt-uppercase _ls-0--44px" data-slot="text">Previous</span></div><span class="is_Text _col-text-primar121 _fs-f-size-3 _lh-f-lineHeigh112922 _fw-500" data-slot="text">Sandbox cold start</span></a><a class="is_Text _col-muted-foreg108704142 _pt-6px _pb-6px _pr-12px _pl-12px _btlr-4px _btrr-4px _bbrr-4px _bblr-4px _bg-edge _btw-1px _brw-1px _borderBottomWidth-1px _borderLeftWidth-1px _btc-border _brc-border _borderBottomColor-border _borderLeftColor-border _fs-f-size-2 _lh-f-lineHeigh112922 _bbs-solid _bts-solid _bls-solid _brs-solid" data-slot="text" href="/benchmarks">All Benchmarks Board</a><a class="is_View _fd-column _gap-2px _items-flex-end" href="/benchmarks/gpqa"><div class="is_View _fd-row _items-center _gap-4px"><span class="is_Text _col-muted-foreg108704142 _fs-f-size-1 _lh-f-lineHeigh112922 _tt-uppercase _ls-0--44px" data-slot="text">Next</span><svg aria-hidden="true" data-i18n-mirror="true" fill="none" height="12" stroke="var(--muted-foreground)" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" style="color:var(--muted-foreground);flex-shrink:0" viewbox="0 0 24 24" width="12"><path d="M5 12h14" stroke="var(--muted-foreground)"></path><path d="m12 5 7 7-7 7" stroke="var(--muted-foreground)"></path></svg></div><span class="is_Text _col-text-primar121 _fs-f-size-3 _lh-f-lineHeigh112922 _fw-500" data-slot="text">GPQA-Diamond</span></a></div><div class="is_View _fd-column"><span class="is_Text font_mono _col-muted-foreg108704142 _dsp-block _mb-8px _ff-f-family _fs-f-size-2 _lh-f-lineHeigh112922 _tt-uppercase _ls-0--52px" data-slot="text">Other benchmarks</span><p class="is_Text _col-text-second96872 _dsp-block _maxW-720px _lh-1--3563rem _mt-0px _mr-0px _mb-0px _ml-0px _fs-f-size-2" data-slot="text"><span class="is_Text _col-inherit" data-slot="text"><a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/locomo">LoCoMo</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/memoryagentbench">MemoryAgentBench</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/longmemeval">LongMemEval</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/cronquestions">CronQuestions</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/repobench-r">RepoBench-R</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/locomo-subject-scope">LoCoMo · subject scope</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/locomo-conv">LoCoMo-Conv</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/fleet">Fleet residency</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/goroutine">Live agent footprint</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/sandbox">Sandbox cold start</a></span><span class="is_Text _col-inherit" data-slot="text"> · <a class="is_Text _col-text-second96872" data-slot="text" href="/benchmarks/gpqa">GPQA-Diamond</a></span></p></div></div></div><p class="is_Text font_mono _col-muted-foreg108704142 _dsp-block _mb-10px _ff-f-family _fs-f-size-1 _lh-f-lineHeigh112922" data-slot="text">the committed runs, in <a href="https://github.com/hanzoai/engine/tree/bench/dossier/bench-runs">bench-runs</a>:<!-- --> <span>CUDA<!-- --> <!-- -->Zen-1.7B<!-- --> <!-- -->Q4_K_M</span><span> · <!-- -->CUDA<!-- --> <!-- -->Zen-4B<!-- --> <!-- -->Q4_K_M</span></p></main><!--$--><!--/$--></div>