PageSourceSearch

https://infinity.inc/research

html infinity.inc collected 2026-10-05 05:47:29 UTC 19,946 bytes, 55 lines download raw bytes

1<!doctype html>
2<html lang="en">
3  <head>
4    <meta charset="UTF-8" />
5    <link rel="icon" type="image/png" href="/infinity_small.png" />
6    <link rel="apple-touch-icon" href="/infinity_small.png" />
7    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
8
9    <!--
10      Per-route OG / Twitter meta tags, injected server-side by Caddy's
11      `templates` directive (see Caddyfile) from .OriginalReq.URL.Path, which
12      survives the SPA rewrite to /index.html.
13
14      Each Go-template if/else directive below is wrapped in its own HTML
15      comment. Caddy still evaluates the directives in production. In Vite's dev
16      server the file is served verbatim, and the comment wrapping keeps the
17      template markers from leaking onto the page: an unwrapped directive is a
18      non-whitespace text node in the head, which the HTML parser moves into the
19      body and renders as raw text. In dev every branch's tags land in the head
20      (harmless); production emits a single branch.
21
22      NOTE: this comment must not contain literal template markers — Caddy's
23      template engine parses them even inside HTML comments.
24    -->
25    <title>Research &amp; Technical Deep Dives — Infinity</title>
26    <meta
27      name="description"
28      content="Research and technical deep dives from Infinity: generated inference stacks, autonomous kernel optimization, and running frontier models on new silicon."
29    />
30    <meta property="og:type" content="website" />
31    <meta property="og:site_name" content="Infinity" />
32    <meta property="og:url" content="https://infinity.inc/research" />
33    <meta property="og:title" content="Research &amp; Technical Deep Dives — Infinity" />
34    <meta
35      property="og:description"
36      content="Research and technical deep dives from Infinity: generated inference stacks, autonomous kernel optimization, and running frontier models on new silicon."
37    />
38    <meta property="og:image" content="https://infinity.inc/infinity-logo-padded.png" />
39    <meta property="og:image:alt" content="Infinity" />
40    <meta name="twitter:card" content="summary_large_image" />
41    <meta name="twitter:title" content="Research &amp; Technical Deep Dives — Infinity" />
42    <meta
43      name="twitter:description"
44      content="Research and technical deep dives from Infinity: generated inference stacks, autonomous kernel optimization, and running frontier models on new silicon."
45    />
46    <meta name="twitter:image" content="https://infinity.inc/infinity-logo-padded.png" />
47    
47<script type="module" crossorigin src="/assets/index-CDTwQWFR.js"></script>
47
48    <link rel="stylesheet" crossorigin href="/assets/index-DwZ2dobo.css">
49  <link rel="canonical" href="https://infinity.inc/research" />
50<link rel="stylesheet" href="/research/styles.css" />
51</head>
52  <body>
53    <div id="root"><div class="MuiBox-root css-1x7tk1l"><div class="MuiBox-root css-1pw7xgc"><header class="MuiBox-root css-9pbaq1"><div class="MuiContainer-root MuiContainer-maxWidthLg css-fa5g8i-MuiContainer-root"><div class="MuiBox-root css-ry61am"><div class="MuiBox-root css-0"></div><div class="MuiStack-root css-9fkgbk-MuiStack-root"><a class="MuiButtonBase-root MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit css-jg2qz6-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/#home">Home</a><a class="MuiButtonBase-root MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit css-jg2qz6-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/partnerships/hardware">How It Works</a><a class="MuiButtonBase-root MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit css-jg2qz6-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/partnerships">Partnerships</a><a class="MuiButtonBase-root MuiButton-root MuiButton-text MuiButton-textPrimary MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorPrimary MuiButton-root MuiButton-text MuiButton-textPrimary MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorPrimary css-qhojl5-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/research">Research</a><div class="MuiBox-root css-0"><button class="MuiButtonBase-root MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit MuiButton-root MuiButton-text MuiButton-textInherit MuiButton-sizeSmall MuiButton-textSizeSmall MuiButton-colorInherit css-jg2qz6-MuiButtonBase-root-MuiButton-root" tabindex="0" type="button" aria-haspopup="menu">Company<span class="MuiButton-icon MuiButton-endIcon MuiButton-iconSizeSmall css-jcxoq4-MuiButton-endIcon"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-1cr4oka-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="KeyboardArrowDownIcon"><path d="M7.41 8.59 12 13.17l4.59-4.58L18 10l-6 6-6-6z"></path></svg></span></button></div></div><div class="MuiBox-root css-1bvc4cc"><a class="MuiButtonBase-root MuiButton-root MuiButton-contained MuiButton-containedPrimary MuiButton-sizeSmall MuiButton-containedSizeSmall MuiButton-colorPrimary MuiButton-root MuiButton-contained MuiButton-containedPrimary MuiButton-sizeSmall MuiButton-containedSizeSmall MuiButton-colorPrimary css-1j6h5k1-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/contact/hardware">Become a Partner</a></div></div><div class="MuiStack-root css-1xenpon-MuiStack-root"><a class="MuiButtonBase-root MuiButton-root MuiButton-contained MuiButton-containedPrimary MuiButton-sizeSmall MuiButton-containedSizeSmall MuiButton-colorPrimary MuiButton-root MuiButton-contained MuiButton-containedPrimary MuiButton-sizeSmall MuiButton-containedSizeSmall MuiButton-colorPrimary css-1f23mor-MuiButtonBase-root-MuiButton-root" tabindex="0" href="/contact/hardware">Become a Partner</a><button class="MuiButtonBase-root MuiIconButton-root MuiIconButton-sizeMedium css-78trlr-MuiButtonBase-root-MuiIconButton-root" tabindex="0" type="button" aria-label="Open menu" aria-haspopup="dialog"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-i4bv87-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="MenuIcon"><path d="M3 18h18v-2H3zm0-5h18v-2H3zm0-7v2h18V6z"></path></svg></button></div></div></header><div class="MuiContainer-root MuiContainer-maxWidthLg css-1tgolqw-MuiContainer-root"><div class="MuiStack-root css-zqvkf0-MuiStack-root"><h4 class="MuiTypography-root MuiTypography-h4 css-nwjqys-MuiTypography-root">Research papers and technical deep-dives from the Infinity team.</h4></div></div><div class="MuiContainer-root MuiContainer-maxWidthLg css-1jep6rt-MuiContainer-root"><div class="MuiBox-root css-iq7471"><a class="MuiBox-root css-1vraa1x" href="/research/dmatrix-full-model-inference"><div class="MuiBox-root css-bx1dzy"><div class="MuiBox-root css-2t7nkg"><picture class="MuiBox-root css-4rx3j2">
53<source media="(max-width: 599.95px)" srcSet="/research/dmatrix-full-model-inference-card-mobile.png"/><img class="MuiBox-root css-41tybk" src="/research/dmatrix-full-model-inference-card.png" alt=""/></picture></div><div class="MuiBox-root css-em3h8m"></div></div><div class="MuiBox-root css-o8e28y"><div class="MuiStack-root css-1vjnajo-MuiStack-root"><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-n4cf5b-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Technical Deep Dive</span></div></div><h5 class="MuiTypography-root MuiTypography-h5 css-vvg93t-MuiTypography-root">Performant Full-Model Inference on d-Matrix Corsair: 20x Speedup</h5><p class="MuiTypography-root MuiTypography-body2 css-1csf2r7-MuiTypography-root">Infinity and d-Matrix worked hand in hand to optimize the performance of Qwen 3 by 20x tokens / second on a single Corsair card compared with Infinity’s original implementation. Context length is also extended by 16x.</p><div class="MuiStack-root css-n9emwe-MuiStack-root"><div class="MuiStack-root css-193z64x-MuiStack-root"><span class="MuiTypography-root MuiTypography-caption css-1jma7xx-MuiTypography-root">July 2026</span></div><div class="MuiStack-root css-a28hco-MuiStack-root"><div class="MuiBox-root css-1iwpysi"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-hwaqf3-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="NorthEastIcon"><path d="M9 5v2h6.59L4 18.59 5.41 20 17 8.41V15h2V5z"></path></svg></div><p class="MuiTypography-root MuiTypography-body1 css-twp891-MuiTypography-root">Read deep dive</p></div></div></div></a><a class="MuiBox-root css-1vraa1x" href="/research/dmatrix-partnership"><div class="MuiBox-root css-1pgom13"><div class="MuiBox-root css-2t7nkg"><picture class="MuiBox-root css-4rx3j2"><img class="MuiBox-root css-41tybk" src="/research/dmatrix-partnership-card.png" alt=""/></picture></div><div class="MuiBox-root css-em3h8m"></div></div><div class="MuiBox-root css-o8e28y"><div class="MuiStack-root css-1vjnajo-MuiStack-root"><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-n4cf5b-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Partnership Announcement</span></div></div><h5 class="MuiTypography-root MuiTypography-h5 css-vvg93t-MuiTypography-root">Infinity × d-Matrix: Partnering to Advance Performant Full-Model Inference</h5><p class="MuiTypography-root MuiTypography-body2 css-1csf2r7-MuiTypography-root">Infinity and d-Matrix have partnered to enable and optimize performant full-model inference on d-Matrix Corsair — bringing Qwen3 from its first tensor-parallel matrix operations to complete, stateful inference on a single Corsair card.</p><div class="MuiStack-root css-n9emwe-MuiStack-root"><div class="MuiStack-root css-193z64x-MuiStack-root"><span class="MuiTypography-root MuiTypography-caption css-1jma7xx-MuiTypography-root">July 2026</span></div><div class="MuiStack-root css-a28hco-MuiStack-root"><div class="MuiBox-root css-1iwpysi"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-hwaqf3-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="NorthEastIcon"><path d="M9 5v2h6.59L4 18.59 5.41 20 17 8.41V15h2V5z"></path></svg></div><p class="MuiTypography-root MuiTypography-body1 css-twp891-MuiTypography-root">Read announcement</p></div></div></div></a><a class="MuiBox-root css-1vraa1x" href="/research/dmatrix-corsair"><div class="MuiBox-root css-1pgom13"><div class="MuiBox-root css-2t7nkg"><picture class="MuiBox-root css-4rx3j2"><img class="MuiBox-root css-41tybk" src="/research/dmatrix-corsair-chip.png" alt=""/></picture></div><div class="MuiBox-root css-em3h8m"></div></div><div class="MuiBox-root css-o8e28y"><div class="MuiStack-root css-1vjnajo-MuiStack-root"><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-n4cf5b-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Announcement</span></div></div><h5 class="MuiTypography-root MuiTypography-h5 css-vvg93t-MuiTypography-root">Igniting d-Matrix: LLM Inference on New Silicon, in Days, Not Years</h5><p class="MuiTypography-root MuiTypography-body2 css-1csf2r7-MuiTypography-root">Within 10 hours we had matrix multiplications at 90%+ of the chip’s theoretical peak. Within 10 days, Qwen3 was running end-to-e
53nd—every operation written from scratch.</p><div class="MuiStack-root css-n9emwe-MuiStack-root"><div class="MuiStack-root css-193z64x-MuiStack-root"><span class="MuiTypography-root MuiTypography-caption css-1jma7xx-MuiTypography-root">April 2026</span></div><div class="MuiStack-root css-a28hco-MuiStack-root"><div class="MuiBox-root css-1iwpysi"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-hwaqf3-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="NorthEastIcon"><path d="M9 5v2h6.59L4 18.59 5.41 20 17 8.41V15h2V5z"></path></svg></div><p class="MuiTypography-root MuiTypography-body1 css-twp891-MuiTypography-root">Read announcement</p></div></div></div></a><a class="MuiBox-root css-1vraa1x" href="/research/qwen3-optimization"><div class="MuiBox-root css-bx1dzy"><div class="MuiBox-root css-2t7nkg"><picture class="MuiBox-root css-4rx3j2"><img class="MuiBox-root css-41tybk" src="/research/qwen3-vllm-comparison.png" alt=""/></picture></div><div class="MuiBox-root css-em3h8m"></div></div><div class="MuiBox-root css-o8e28y"><div class="MuiStack-root css-1vjnajo-MuiStack-root"><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-n4cf5b-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Case Study</span></div><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-1fb4h1p-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Hacker News front page</span></div></div><h5 class="MuiTypography-root MuiTypography-h5 css-vvg93t-MuiTypography-root">Surpassing vLLM with a Generated Inference Stack</h5><p class="MuiTypography-root MuiTypography-body2 css-1csf2r7-MuiTypography-root">Infinity&#x27;s infy optimization system wrote an inference engine from scratch and autonomously optimized it on Qwen3-8B. The resulting engine delivers up to 34.3% more tokens per second than vLLM when configured with identical parameters.</p><div class="MuiStack-root css-n9emwe-MuiStack-root"><div class="MuiStack-root css-193z64x-MuiStack-root"><span class="MuiTypography-root MuiTypography-caption css-1jma7xx-MuiTypography-root">March 2026</span></div><div class="MuiStack-root css-a28hco-MuiStack-root"><div class="MuiBox-root css-1iwpysi"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-hwaqf3-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="NorthEastIcon"><path d="M9 5v2h6.59L4 18.59 5.41 20 17 8.41V15h2V5z"></path></svg></div><p class="MuiTypography-root MuiTypography-body1 css-twp891-MuiTypography-root">Read research</p></div></div></div></a><a class="MuiBox-root css-1vraa1x" href="/research/OMEGA%20-%20ICLR%20Paper.pdf" target="_blank" rel="noreferrer"><div class="MuiBox-root css-1pgom13"><div class="MuiBox-root css-2t7nkg"><picture class="MuiBox-root css-4rx3j2"><img class="MuiBox-root css-41tybk" src="/research/omega-framework.png" alt=""/></picture></div><div class="MuiBox-root css-em3h8m"></div></div><div class="MuiBox-root css-o8e28y"><div class="MuiStack-root css-1vjnajo-MuiStack-root"><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-n4cf5b-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">Research Paper</span></div><div class="MuiChip-root MuiChip-filled MuiChip-sizeSmall MuiChip-colorDefault MuiChip-filledDefault css-1fb4h1p-MuiChip-root"><span class="MuiChip-label MuiChip-labelSmall css-wjsjww-MuiChip-label">ICLR 2026</span></div></div><h5 class="MuiTypography-root MuiTypography-h5 css-vvg93t-MuiTypography-root">OMEGA: Optimizing Machine Learning by Evaluating Generated Algorithms</h5><p class="MuiTypography-root MuiTypography-body2 css-2ta5aa-MuiTypography-root">In order to automate AI research we introduce a full, end-to-end framework, OMEGA: Optimizing Machine learning by Evaluating Generated Algorithms, that starts at idea generation and ends with executable code. Our system combines structured meta-prompt engineering with executable code generation to create new ML classifiers. The OMEGA framework has been utilized to generate several novel algorithms that outperform scikit-learn baselines across a robust selection of 20 benchmark datasets (infinity-bench). You can access models discussed in this paper and more in the python package: pip install omega-models.</p><div class="MuiStack-root css-n9emwe-MuiStack-root"><div class="MuiStack-root css-193z64x-MuiStack-root">
53<span class="MuiTypography-root MuiTypography-caption css-1jma7xx-MuiTypography-root">March 2026</span></div><div class="MuiStack-root css-a28hco-MuiStack-root"><div class="MuiBox-root css-1iwpysi"><svg class="MuiSvgIcon-root MuiSvgIcon-fontSizeMedium css-hwaqf3-MuiSvgIcon-root" focusable="false" aria-hidden="true" viewBox="0 0 24 24" data-testid="DescriptionOutlinedIcon"><path d="M8 16h8v2H8zm0-4h8v2H8zm6-10H6c-1.1 0-2 .9-2 2v16c0 1.1.89 2 1.99 2H18c1.1 0 2-.9 2-2V8zm4 18H6V4h7v5h5z"></path></svg></div><p class="MuiTypography-root MuiTypography-body1 css-twp891-MuiTypography-root">View paper</p></div></div></div></a></div></div></div><footer class="MuiBox-root css-fvhti5"><div class="MuiContainer-root MuiContainer-maxWidthLg css-kojkh3-MuiContainer-root"><div class="MuiBox-root css-1g4lsp"><div class="MuiStack-root css-hx5qeh-MuiStack-root"><img class="MuiBox-root css-1kyycar" src="/infinity.png" alt="Infinity"/><p class="MuiTypography-root MuiTypography-body2 css-3d5j3i-MuiTypography-root">Model-aware inference software that makes any AI chip production-ready.</p><div class="MuiStack-root css-5u9cht-MuiStack-root"><a class="MuiBox-root css-zhl842" href="https://x.com/infinity_ai_" target="_blank" rel="noreferrer" aria-label="X"><svg class="MuiBox-root css-ola6kf" viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z"></path></svg></a><a class="MuiBox-root css-zhl842" href="https://www.linkedin.com/company/infinity-ai-institute" target="_blank" rel="noreferrer" aria-label="LinkedIn"><svg class="MuiBox-root css-ola6kf" viewBox="0 0 24 24"><path d="M20.447 20.452h-3.554v-5.569c0-1.328-.027-3.037-1.852-3.037-1.853 0-2.136 1.445-2.136 2.939v5.667H9.351V9h3.414v1.561h.046c.477-.9 1.637-1.85 3.37-1.85 3.601 0 4.267 2.37 4.267 5.455v6.286zM5.337 7.433a2.062 2.062 0 01-2.063-2.065 2.064 2.064 0 112.063 2.065zm1.782 13.019H3.555V9h3.564v11.452zM22.225 0H1.771C.792 0 0 .774 0 1.729v20.542C0 23.227.792 24 1.771 24h20.451C23.2 24 24 23.227 24 22.271V1.729C24 .774 23.2 0 22.225 0z"></path></svg></a></div></div><div class="MuiStack-root css-j3ygch-MuiStack-root"><span class="MuiTypography-root MuiTypography-overline css-c1tjpe-MuiTypography-root">Company</span><a class="MuiBox-root css-6w481y" href="/team">Team</a><a class="MuiBox-root css-6w481y" href="/careers">Careers</a><a class="MuiBox-root css-6w481y" href="/partnerships">Partnerships</a><a class="MuiBox-root css-6w481y" href="/contact/hardware">Become a Partner</a></div><div class="MuiStack-root css-j3ygch-MuiStack-root"><span class="MuiTypography-root MuiTypography-overline css-c1tjpe-MuiTypography-root">News</span><a class="MuiBox-root css-6w481y" href="/press">Newsroom</a><a class="MuiBox-root css-6w481y" href="/press#in-the-news">In the News</a><a class="MuiBox-root css-6w481y" href="/press#press-kit">Press Kit</a></div><div class="MuiStack-root css-j3ygch-MuiStack-root"><span class="MuiTypography-root MuiTypography-overline css-c1tjpe-MuiTypography-root">Insights</span><a class="MuiBox-root css-6w481y" href="/research">Research</a></div></div><hr class="MuiDivider-root MuiDivider-fullWidth css-3mxnf-MuiDivider-root"/><div class="MuiStack-root css-w8s52o-MuiStack-root"><span class="MuiTypography-root MuiTypography-caption css-epje7o-MuiTypography-root">© Infinity Inc. · infinity.inc · 2026</span></div></div></footer></div></div>
54  </body>
55</html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.