Download index.html from HuggingFaceBio/Carbon-X-Release: direct link, hf CLI and curl.
- Browser
- Download file 66.4 kB
-
https://huggingface.co/spaces/HuggingFaceBio/Carbon-X-Release/resolve/main/index.html
- Command line
-
hf download hf://spaces/HuggingFaceBio/Carbon-X-Release/index.html
-
curl -L -o index.html https://huggingface.co/spaces/HuggingFaceBio/Carbon-X-Release/resolve/main/index.html
66.4 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1"> | |
| <title>Carbon-X: When Natural Language Meets DNA</title> | |
| <link rel="preconnect" href="https://fonts.googleapis.com"> | |
| <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin> | |
| <link href="https://fonts.googleapis.com/css2?family=Newsreader:opsz,wght@6..72,400;6..72,500&family=Inter:wght@400;500;600&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet"> | |
| <style> | |
| /* Design tokens. Settings chosen in the design mockups: warm-white paper, full green, 15px body, 960px column. */ | |
| :root { | |
| --paper: #fefdfc; | |
| --rule: #e7e6e4; | |
| --ink: #1c1b18; | |
| --ink-2: #4a4843; | |
| --ink-3: #85827a; | |
| --green: #008225; | |
| --green-ink: #006729; | |
| --tbd-bg: rgba(192, 137, 47, 0.16); | |
| --tbd-ink: #7a5a1c; | |
| --measure: 960px; | |
| --fs-title: 46px; --fs-h2: 27px; --fs-lede: 20.5px; --fs-body: 15px; | |
| } | |
| @media (prefers-color-scheme: dark) { | |
| :root { | |
| --paper: #17171a; --rule: #2e2e31; --ink: #ecebe6; --ink-2: #bdbbb4; --ink-3: #8c8a83; | |
| --green: #5fb27a; --green-ink: #7cc493; --tbd-bg: rgba(214, 165, 80, 0.18); --tbd-ink: #e0b66a; | |
| } | |
| } | |
| * { box-sizing: border-box; margin: 0; padding: 0; } | |
| html { scroll-behavior: smooth; } | |
| body { background: var(--paper); color: var(--ink); font-family: Inter, "Helvetica Neue", system-ui, sans-serif; | |
| font-size: var(--fs-body); line-height: 1.72; -webkit-font-smoothing: antialiased; } | |
| .col { max-width: var(--measure); margin: 0 auto; padding: 0 20px; } | |
| /* header: serif title and standfirst, everything else in Inter */ | |
| header.col { padding-top: 96px; } | |
| h1 { font-family: Newsreader, Georgia, serif; font-weight: 400; font-size: var(--fs-title); line-height: 1.1; | |
| letter-spacing: -0.02em; margin-bottom: 28px; text-wrap: balance; } | |
| .lede { font-family: Newsreader, Georgia, serif; font-size: var(--fs-lede); line-height: 1.4; color: var(--ink-2); margin-bottom: 30px; text-wrap: pretty; } | |
| .standfirst { color: var(--ink-2); margin-bottom: 30px; text-wrap: pretty; } | |
| .partner-logos { display: grid; grid-template-columns: 1.2fr 1.4fr .9fr 1.3fr; align-items: center; gap: 28px; margin-top: 28px; padding: 24px; background: #fff; border-radius: 8px; } | |
| .partner-logos a { display: flex; align-items: center; justify-content: center; min-width: 0; } | |
| .partner-logos img, .partner-logos svg { display: block; width: 100%; height: 72px; object-fit: contain; } | |
| .partner-logos a:focus-visible { outline: 2px solid var(--green); outline-offset: 6px; } | |
| @media (max-width: 600px) { .partner-logos { grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 24px; padding: 20px; } } | |
| .byline { display: flex; flex-direction: column; gap: 10px; } | |
| .authors { display: flex; flex-wrap: wrap; align-items: center; gap: 8px 18px; } | |
| .author { display: inline-flex; align-items: center; gap: 8px; color: var(--ink) ; font-weight: 500; } | |
| .author img { width: 26px; height: 26px; border-radius: 50%; object-fit: cover; background: var(--rule); } | |
| .release { display: flex; flex-wrap: wrap; align-items: center; gap: 4px 8px; } | |
| .release .label { color: var(--ink-3); } | |
| .links { padding: 14px 0; border-top: 1px solid var(--rule); border-bottom: 1px solid var(--rule); font-size: 14px; margin-bottom: 32px; } | |
| .links a { color: var(--green-ink); font-weight: 500; text-decoration: none; } | |
| .links a:hover { text-decoration: underline; text-underline-offset: 3px; } | |
| .hl-lead { margin-bottom: 6px; } | |
| ul.highlights { margin-bottom: 28px; } | |
| ul.highlights li { margin-bottom: 6px; } | |
| /* body text */ | |
| h2 { font-family: Newsreader, Georgia, serif; font-weight: 500; font-size: var(--fs-h2); line-height: 1.2; letter-spacing: -0.01em; | |
| margin: 64px 0 16px; text-wrap: balance; } | |
| h3 { font-size: 1.1em; font-weight: 600; margin: 28px 0 8px; } | |
| p { margin-bottom: 1.05em; text-wrap: pretty; } | |
| p strong, li b, li strong { font-weight: 600; } | |
| p.soft { color: var(--ink-2); } | |
| a.inline { color: var(--green-ink); text-decoration: underline; text-decoration-thickness: 1px; text-underline-offset: 3px; } | |
| code { font-family: "JetBrains Mono", monospace; font-size: 0.85em; background: rgba(0,0,0,.04); padding: 1px 5px; border-radius: 4px; } | |
| ul.clean { list-style: none; margin: 4px 0 18px; } | |
| ul.clean li { position: relative; padding-left: 20px; margin-bottom: 8px; } | |
| ul.clean li::before { content: ""; position: absolute; left: 4px; top: 0.72em; width: 5px; height: 5px; border-radius: 50%; background: var(--green); } | |
| ul.clean ul.clean { margin: 8px 0 4px; } | |
| ul.clean ul.clean li::before { background: transparent; border: 1.5px solid var(--green); width: 4px; height: 4px; } | |
| .tbd { background: var(--tbd-bg); color: var(--tbd-ink); font: 500 .8em/1.4 "JetBrains Mono", monospace; padding: 1px 5px; border-radius: 3px; | |
| -webkit-box-decoration-break: clone; box-decoration-break: clone; } | |
| /* figures */ | |
| figure { margin: 48px 0; } | |
| .placeholder { border: 1px dashed var(--ink-3); border-radius: 4px; padding: 36px 24px; text-align: center; color: var(--ink-3); font-size: 14px; line-height: 1.5; } | |
| .placeholder.tall { padding: 110px 24px; } | |
| .placeholder .ph-label { display: block; font: 500 11px "JetBrains Mono", monospace; letter-spacing: .1em; text-transform: uppercase; color: var(--tbd-ink); margin-bottom: 6px; } | |
| figure img, figure video { display: block; width: 100%; height: auto; border-radius: 4px; } | |
| .figtitle { font-size: 16px; font-weight: 600; line-height: 1.35; margin-bottom: 14px; } | |
| figcaption { font-size: 13.5px; line-height: 1.55; color: var(--ink-2); margin-top: 12px; } | |
| figcaption b { color: var(--ink); font-weight: 600; } | |
| /* TMP watermark: marks figures that are screenshots from the draft and still need to be rebuilt as native HTML */ | |
| .tmp { position: relative; overflow: hidden; } | |
| .tmp::after { content: "TMP"; position: absolute; inset: 0; display: flex; align-items: center; justify-content: center; | |
| font: 800 clamp(56px, 14vw, 180px)/1 Inter, sans-serif; letter-spacing: .12em; color: rgba(200, 30, 30, .16); | |
| transform: rotate(-16deg); pointer-events: none; } | |
| .chart { position: relative; } | |
| .chart svg { display: block; width: 100%; height: auto; overflow: visible; } | |
| .chart svg text { font-family: Inter, sans-serif; } | |
| .chart .pt { outline: none; } | |
| .chart .pt:hover .logo, .chart .pt:focus .logo { filter: drop-shadow(0 0 3px var(--ink-3)); } | |
| /* benchmark score table (Figure 2) */ | |
| .scores { overflow-x: auto; } | |
| .scores table { min-width: 760px; table-layout: fixed; font-size: 14px; font-variant-numeric: normal; } | |
| .scores th, .scores td { padding: 12px 8px; text-align: center; border-bottom: 1px solid var(--rule); } | |
| .scores thead th { vertical-align: bottom; font-size: 13px; font-weight: 600; color: var(--ink); white-space: normal; line-height: 1.3; border-bottom: 1px solid var(--ink-3); } | |
| .scores thead th:first-child, .scores tbody th { width: 27%; } | |
| .scores tbody tr { height: 62px; } /* every row, including Average, the height of a two-line benchmark label */ | |
| .scores .effort { display: block; font-size: 11.5px; font-weight: 400; color: var(--ink-3); margin-top: 6px; } | |
| .scores tbody th { vertical-align: middle; text-align: left; font-size: 14px; font-weight: 500; color: var(--ink); white-space: normal; line-height: 1.35; position: sticky; left: 0; background: var(--paper); } | |
| .scores .bench { display: block; font-size: 12px; font-weight: 400; color: var(--ink-3); margin-top: 2px; } | |
| .scores tbody td { font-size: 14.5px; color: var(--ink-2); font-variant-numeric: tabular-nums; vertical-align: middle; } | |
| .scores .carbon { background: color-mix(in srgb, var(--green) 8%, var(--paper)); color: var(--green-ink); font-weight: 700; } | |
| .scores thead .carbon { color: var(--green-ink); } | |
| .scores .best { background: color-mix(in srgb, var(--ink) 7%, var(--paper)); font-weight: 700; color: var(--ink); } | |
| .scores .carbon.best { background: color-mix(in srgb, var(--green) 16%, var(--paper)); color: var(--green-ink); } | |
| .scores tr.avg th, .scores tr.avg td { border-top: 1px solid var(--ink-3); } | |
| .scores tr.avg th { font-weight: 600; } | |
| /* chart legends */ | |
| .legend { display: flex; flex-wrap: wrap; gap: 6px 18px; font-size: 13px; color: var(--ink-2); margin-bottom: 10px; } | |
| .legend i { display: inline-block; width: 14px; height: 3px; border-radius: 2px; margin-right: 7px; vertical-align: 4px; } | |
| .legend .sep { margin-left: auto; } | |
| .legend b { display: inline-block; width: 10px; height: 10px; border-radius: 50%; margin-right: 6px; vertical-align: -1px; border: 2px solid var(--ink-2); } | |
| .legend b.dot { background: var(--ink-2); } | |
| /* embedded figures (figures/*.html) */ | |
| iframe.fig { display: block; width: 100%; height: 420px; border: 0; background: transparent; } | |
| /* collapsible boxes inside the recipe */ | |
| details.box { border: 1px solid var(--rule); border-radius: 6px; padding: 0 16px; margin: 24px 0; } | |
| details.box:last-of-type { border-bottom: 1px solid var(--rule); } | |
| details.box summary { font-weight: 600; padding: 14px 30px 14px 0; } | |
| details.box summary::after { top: 10px; } | |
| details.box[open] summary { border-bottom: 1px solid var(--rule); margin-bottom: 14px; } | |
| h3 .tbd { font-size: 11px; vertical-align: 3px; } | |
| /* Figure 2: Carbon + Qwen -> Carbon-X lineage diagram */ | |
| .lineage { display: grid; gap: 14px; } | |
| .lineage .row { display: grid; grid-template-columns: minmax(0, 1fr) auto 120px auto minmax(0, 1fr); align-items: center; gap: 12px; } | |
| .lineage .io { display: flex; flex-wrap: wrap; gap: 5px; } | |
| .lineage .row .io:first-child { justify-content: flex-end; } | |
| .lineage .arr { color: var(--ink-3); } | |
| .lineage .model { height: 38px; display: flex; align-items: center; justify-content: center; border: 1px solid var(--ink-3); border-radius: 6px; font-weight: 600; font-size: 14px; background: var(--paper); } | |
| .lineage .cx .model { border: 2px solid var(--ink); } | |
| .tk { font: 500 12px/1 "JetBrains Mono", monospace; padding: 6px 7px; border-radius: 4px; white-space: nowrap; } | |
| .tk.dna { background: color-mix(in srgb, #c0892f 20%, var(--paper)); color: #7a5a1c; } | |
| .tk.txt { background: color-mix(in srgb, var(--green) 13%, var(--paper)); color: var(--green-ink); } | |
| @media (max-width: 760px) { | |
| .lineage .row { grid-template-columns: minmax(0, 1fr) auto 78px auto minmax(0, 1fr); gap: 6px; } | |
| .lineage .model { height: 32px; font-size: 12.5px; } | |
| .tk { font-size: 10.5px; padding: 4px 5px; } | |
| } | |
| /* Figure 3: architecture (left to right; the token sequence runs top to bottom) */ | |
| .arch-wrap { container-type: inline-size; } | |
| .arch { --gold: #c0892f; --gold-ink: #7a5a1c; --row: 24px; --rgap: 4px; --ggap: 12px; | |
| display: flex; align-items: center; justify-content: space-between; gap: 6px; min-width: 0; } | |
| .arch .arr { color: var(--ink-3); font-size: 14px; flex: none; } | |
| .arch .panel { border: 1px solid var(--rule); border-radius: 8px; padding: 10px 8px 12px; background: var(--paper); flex: none; } | |
| .arch .panel h4 { font-size: 12px; font-weight: 600; line-height: 1.3; margin-bottom: 10px; white-space: nowrap; } | |
| .arch .panel.core { border: 1.5px solid var(--ink); } | |
| .arch .field { font: 400 11.5px/1.55 "JetBrains Mono", monospace; color: var(--ink-2); border: 1px solid var(--rule); border-radius: 5px; padding: 8px 9px; background: #fff; white-space: nowrap; } | |
| .arch .field .seq { display: block; color: var(--gold-ink); } | |
| /* a group = a vertical label next to a column of rows */ | |
| .arch .groups { display: flex; flex-direction: column; gap: var(--ggap); } | |
| .arch .group { display: flex; gap: 7px; align-items: stretch; } | |
| .arch .vlab { writing-mode: vertical-rl; transform: rotate(180deg); font-size: 10.5px; font-weight: 600; text-align: center; white-space: nowrap; | |
| padding-left: 5px; border-left: 1.5px solid currentColor; } | |
| .arch .vlab.txt { color: var(--green-ink); } .vlab.dna { color: var(--gold-ink); } | |
| .arch .rows { display: flex; flex-direction: column; gap: var(--rgap); } | |
| .arch .tokrows { width: 84px; } /* text and DNA tokens share one width */ | |
| .arch .tk { height: var(--row); display: flex; align-items: center; font: 500 11px/1 "JetBrains Mono", monospace; padding: 0 7px; border-radius: 4px; white-space: nowrap; } | |
| .arch .tk.txt { background: color-mix(in srgb, var(--green) 14%, var(--paper)); color: var(--green-ink); } | |
| .arch .tk.dna { background: color-mix(in srgb, var(--gold) 22%, var(--paper)); color: var(--gold-ink); } | |
| /* one embedding vector per token, drawn as a strip of cells */ | |
| .arch .vec { height: var(--row); display: flex; gap: 2px; align-items: stretch; padding: 3px 0; } | |
| .arch .vec i { width: 9px; border-radius: 1.5px; } | |
| .arch .vec.txt i { background: color-mix(in srgb, var(--green) var(--s), var(--paper)); } | |
| .arch .vec.dna i { background: color-mix(in srgb, var(--gold) var(--s), var(--paper)); } | |
| @container (max-width: 850px) { | |
| .arch { flex-direction: column; gap: 12px; } | |
| .arch > .arr { transform: rotate(90deg); } | |
| } | |
| /* backbone: layers side by side, each spanning the whole sequence */ | |
| .arch .layers { display: flex; gap: 5px; height: calc(8 * var(--row) + 7 * var(--rgap) + var(--ggap)); } | |
| .arch .layers div { writing-mode: vertical-rl; transform: rotate(180deg); display: flex; align-items: center; justify-content: center; width: 30px; | |
| font-size: 11px; color: var(--ink-2); border-radius: 4px; background: color-mix(in srgb, var(--ink) 6%, var(--paper)); white-space: nowrap; } | |
| .arch .layers .gap { background: none; width: 14px; color: var(--ink-3); } | |
| .arch .layers .head { background: color-mix(in srgb, var(--ink) 12%, var(--paper)); color: var(--ink); font-weight: 500; } | |
| /* table */ | |
| .tablewrap { overflow-x: auto; margin: 20px 0 12px; } | |
| table { width: 100%; border-collapse: collapse; font-size: 14px; font-variant-numeric: tabular-nums; } | |
| th { text-align: left; font-weight: 500; font-size: 12px; color: var(--ink-3); padding: 8px 10px; border-bottom: 1px solid var(--ink-3); white-space: nowrap; } | |
| th.num, td.num { text-align: right; } | |
| td { padding: 11px 10px; border-bottom: 1px solid var(--rule); vertical-align: top; } | |
| td .bname { font-weight: 600; display: block; } | |
| td .bdesc { color: var(--ink-3); font-size: 13px; line-height: 1.45; } | |
| td.carbon { color: var(--green-ink); font-weight: 600; } | |
| td .who { display: block; font-size: 12px; color: var(--ink-3); } | |
| /* training recipe steps */ | |
| details { border-top: 1px solid var(--rule); } | |
| details:last-of-type { border-bottom: 1px solid var(--rule); } | |
| summary { list-style: none; cursor: pointer; padding: 18px 34px 18px 0; position: relative; } | |
| summary::-webkit-details-marker { display: none; } | |
| summary::after { content: "+"; position: absolute; right: 4px; top: 16px; font-size: 22px; font-weight: 300; color: var(--ink-3); } | |
| details[open] summary::after { content: "–"; } | |
| .stepno { display: block; font: 500 11px "JetBrains Mono", monospace; letter-spacing: .1em; color: var(--green-ink); margin-bottom: 2px; } | |
| summary .dtitle { display: block; font-family: Newsreader, Georgia, serif; font-size: 20px; font-weight: 500; line-height: 1.3; } | |
| summary .dtake { display: block; color: var(--ink-2); margin-top: 3px; } | |
| details .dbody { padding: 0 0 20px; } | |
| .draftnote { display: inline-block; margin-top: 40px; font: 500 11.5px "JetBrains Mono", monospace; color: var(--tbd-ink); background: var(--tbd-bg); border-radius: 4px; padding: 6px 10px; } | |
| footer { margin-top: 80px; padding: 26px 0 56px; border-top: 1px solid var(--rule); font-size: 13px; color: var(--ink-3); } | |
| @media (max-width: 760px) { | |
| :root { --fs-title: 33px; --fs-h2: 22px; --fs-lede: 18.5px; } | |
| header.col { padding-top: 48px; } | |
| h2 { margin-top: 52px; } | |
| .placeholder.tall { padding: 64px 20px; } | |
| .legend .sep { margin-left: 0; } | |
| } | |
| </style> | |
| </head> | |
| <body> | |
| <header class="col"> | |
| <h1>Carbon-X: When Natural Language Meets DNA</h1> | |
| <p class="standfirst"><em>We midtrained a 14B model on a combination of DNA and natural language data to excel at DNA understanding tasks. | |
| It beats every other open language model and delivers frontier intelligence that matches GPT-6.1 Sol at 1,000× lower cost. | |
| You can ask about a sequence, predict what a mutation does, or generate new DNA. | |
| It’s the first step towards a deeper understanding of DNA with language.</em></p> | |
| <div class="links byline"> | |
| <div class="authors"> | |
| <a class="author" href="https://huggingface.co/loubnabnl" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/61c141342aac764ce1654e43/81AwoT5IQ_Xdw0OVw7TKu.jpeg" alt="" width="26" height="26">Loubna Ben Allal</a> | |
| <a class="author" href="https://huggingface.co/lewtun" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/1594651707950-noauth.jpeg" alt="" width="26" height="26">Lewis Tunstall</a> | |
| <a class="author" href="https://huggingface.co/danaaubakirova" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/640e21ef3c82bd463ee5a76d/nVR1DFPAsiLw6Boys28Rb.jpeg" alt="" width="26" height="26">Dana Aubakirova</a> | |
| <a class="author" href="https://huggingface.co/mauryzio" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/6923888fa13feec3d654d7b3/IATU3iUYAnM2uftp8qtKi.jpeg" alt="" width="26" height="26">Maurizio Fiusco</a> | |
| <a class="author" href="https://huggingface.co/edbeeching" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/1644220542819-noauth.jpeg" alt="" width="26" height="26">Edward Beeching</a> | |
| <a class="author" href="https://huggingface.co/kashif" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/1669189789447-629f3b18ee05727ce328ccbe.jpeg" alt="" width="26" height="26">Kashif Rasul</a> | |
| <a class="author" href="https://huggingface.co/cgeorgiaw" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/WEjkuS_TxIgtYRxNPa0VS.png" alt="" width="26" height="26">Georgia Channing</a> | |
| <a class="author" href="https://huggingface.co/lvwerra" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/5e48005437cb5b49818287a5/4uCXGGui-9QifAT4qelxU.png" alt="" width="26" height="26">Leandro von Werra</a> | |
| <a class="author" href="https://huggingface.co/GenerTeam" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/67aac08cac45ae07ef08d203/1H1lTTlxXGRWqGPCTRmIM.jpeg" alt="" width="26" height="26">Qiuyi Li</a> | |
| <a class="author" href="https://huggingface.co/Diegodibi" target="_blank" rel="noopener noreferrer"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/69130238c99d4150b0a6c8c7/RoUkk0lbWR605s-ASuhhE.jpeg" alt="" width="26" height="26">Diego di Bernardo</a> | |
| </div> | |
| <div class="release"><span class="label">Model:</span><a class="modellink" href="https://huggingface.co/HuggingFaceBio/Carbon-X" target="_blank" rel="noopener noreferrer">huggingface.co/HuggingFaceBio/Carbon-X</a> </div> | |
| </div> | |
| <p>We’re releasing <strong>Carbon-X, an open 14-billion-parameter model for DNA understanding and generation</strong>. Carbon-X brings DNA sequences and natural language into one model: researchers can ask questions about a sequence, predict the effects of mutations, and generate new DNA. It can also act as a specialist sub-agent, helping other models interpret sequences they struggle to understand on their own.</p> | |
| <p>Language models learn about biology from scientific literature, but that knowledge does not necessarily translate into understanding raw DNA. A model may explain what a gene does yet struggle to assess a change in its sequence. We built Carbon-X to connect these capabilities: learning directly from genomes while retaining the ability to answer questions in natural language.</p> | |
| <p>Starting from Qwen3-14B, we trained on 300 billion tokens combining DNA and text, then used supervised fine-tuning and reinforcement learning to teach the model to answer questions about sequences. We’re releasing Carbon-X under Apache 2.0 so researchers can run it on their own hardware and adapt it to their work, keeping sensitive or unpublished sequences under their control.</p> | |
| <p>Across five benchmarks of DNA understanding and generation, Carbon-X outperforms every other open model we evaluated. Its average score approaches GPT-6.1 Sol’s at 1/1,800 of the inference cost per problem. This is a 27-point improvement over Qwen3-14B, showing how much an open language model can gain from training directly on DNA.</p> | |
| <!-- HERO FIGURE: native SVG, drawn by the script at the end of the page --> | |
| <figure class="hero"> | |
| <div class="chart" id="fig1"><noscript><img src="assets/fig1-score-vs-cost.png" width="1135" height="696" alt="Average DNA benchmark score against cost per problem relative to Carbon-X."></noscript></div> | |
| <figcaption><b>Figure 1.</b> Each comparison model is shown at the reasoning effort with its highest intelligence index. Costs are based on standard API pricing for comparison models and measured throughput per H200-hour at $5/hour for Carbon-X.</figcaption> | |
| </figure> | |
| <p><strong>Carbon-X’s strongest results are in predicting mutation effects and completing DNA sequences.</strong> It leads every model we tested, open or closed, on coding-variant effects and next-base prediction. The table below breaks down performance across all five tasks, including where frontier models retain an advantage.</p> | |
| <!-- Data: carbon-x-blog-figures/benchmark-scores.html (leaderboard snapshot 2026-10-21T13:31:46Z). Each model uses the | |
| effort with the highest unrounded mean across the five benchmarks; average is the unweighted mean; best cell per row shaded. --> | |
| <figure> | |
| <div class="scores"><table aria-label="DNA benchmark scores by model, each at its selected reasoning effort"> | |
| <thead><tr><th aria-label="Benchmark"></th><th scope="col" class="carbon">Carbon-X<span class="effort">No reasoning</span></th><th scope="col">GPT-6.1 Sol<span class="effort">Max</span></th><th scope="col">Opus 5.5<span class="effort">Max</span></th><th scope="col">Gemini 4 Argon<span class="effort">Low</span></th><th scope="col">GLM-5.3<span class="effort">Low</span></th><th scope="col">Kimi K3<span class="effort">Low</span></th><th scope="col">DeepSeek V4.1 Flash<span class="effort">Low</span></th></tr></thead> | |
| <tbody> | |
| <tr><th scope="row">Coding variant effects<span class="bench">BioReason Coding</span></th><td class="carbon best">74.9</td><td>71.0</td><td>70.3</td><td>6.5</td><td>52.1</td><td>53.9</td><td>51.4</td></tr> | |
| <tr><th scope="row">Complex variant effects<span class="bench">BioReason Non-SNV</span></th><td class="carbon">85.1</td><td>91.7</td><td class="best">94.8</td><td>90.3</td><td>60.1</td><td>65.3</td><td>55.9</td></tr> | |
| <tr><th scope="row">DNA understanding<span class="bench">GenomeQA</span></th><td class="carbon">63.6</td><td class="best">67.2</td><td>62.2</td><td>63.1</td><td>44.6</td><td>52.0</td><td>49.7</td></tr> | |
| <tr><th scope="row">DNA prediction<span class="bench">NextBase</span></th><td class="carbon best">39.4</td><td>32.9</td><td>1.1</td><td>31.2</td><td>9.7</td><td>30.6</td><td>27.2</td></tr> | |
| <tr><th scope="row">Single-base variant effects<span class="bench">FrontierVEP</span></th><td class="carbon">70.4</td><td>74.9</td><td class="best">76.2</td><td>58.0</td><td>32.7</td><td>50.1</td><td>51.3</td></tr> | |
| <tr class="avg"><th scope="row">Average</th><td class="carbon">66.7</td><td class="best">67.5</td><td>60.9</td><td>49.8</td><td>39.8</td><td>50.4</td><td>47.1</td></tr> | |
| </tbody> | |
| </table></div> | |
| <figcaption><b>Table 1.</b> DNA benchmark scores (%). Each model is shown at the reasoning effort with its highest average score; the best score on each benchmark is shaded.</figcaption> | |
| </figure> | |
| <!-- TO CHECK before publishing: FrontierVEP metric (plain vs balanced accuracy), whether "up to 8,192 bases" | |
| holds for the whole benchmark, and whether the decontamination sentence is still accurate. --> | |
| <details class="box"> | |
| <summary>About the five benchmarks </summary> | |
| <div class="dbody"> | |
| <ul class="clean"> | |
| <li><b>BioReason Coding:</b> compare a reference and a mutated protein-coding sequence and predict whether the mutation is harmful.</li> | |
| <li><b>BioReason Non-SNV:</b> predict whether changes other than single-base swaps, such as insertions and deletions, are benign or harmful.</li> | |
| <li><b>GenomeQA:</b> identify biological features, such as regions that control gene activity, and classify sequences by organism group.</li> | |
| <li><b>NextBase:</b> predict the next bases of a DNA sequence.</li> | |
| <li><b>FrontierVEP:</b> given a long reference sequence (up to 8,192 bases) and a single-base change, predict whether it is benign or pathogenic.</li> | |
| </ul> | |
| <p>BioReason and GenomeQA come from the literature; FrontierVEP and NextBase were built by us. All five are decontaminated against the training data (DNA overlap, text overlap and homology checks).</p> | |
| </div> | |
| </details> | |
| <p>We built Carbon-X by teaching Qwen3-14B DNA as a new modality during midtraining, then post-training it to answer questions about sequences. Below we explain how.</p> | |
| </header> | |
| <main class="col"> | |
| <section> | |
| <h2>Combining DNA and language</h2> | |
| <p>When we train language models on text, we expect them to learn its underlying patterns. At best, they learn its structure as well as the most skilled humans, but we don’t expect them to find patterns unknown to us.</p> | |
| <p>DNA is different: it is not a language we constructed, and our understanding of it is limited. Training models on DNA could let them learn the patterns we already know and uncover new ones in its structure and function.</p> | |
| <p>DNA encodes biological instructions in an alphabet of four letters: A, G, T, and C. Genomic foundation models, such as our earlier <a class="inline" href="https://huggingface.co/HuggingFaceBio/Carbon-3B" target="_blank" rel="noopener noreferrer">Carbon-3B</a>, learn patterns from large collections of DNA sequences, just as language models learn patterns from text. Researchers use them to score sequences, estimate the effects of mutations, or predict which bases come next, but these models cannot explain a prediction or answer a question about a sequence.</p> | |
| <p>Language models trained mainly on text can explain biological concepts, reason about experiments and discuss findings from the literature. But they learn biology largely through descriptions, with little direct exposure to DNA itself. Bringing the two together would let a model connect patterns in DNA with decades of knowledge recorded in papers and databases, so researchers can ask about a specific sequence and discuss what it means.</p> | |
| <figure> | |
| <div class="lineage" role="img" aria-label="Three rows. Carbon takes DNA tokens and outputs DNA tokens. Qwen3 takes text tokens and outputs text tokens. Carbon-X takes mixed text and DNA tokens and outputs mixed text and DNA tokens."> | |
| <div class="row"><div class="io"><span class="tk dna">ATGCGC</span><span class="tk dna">TAGCTA</span><span class="tk dna">ACGTAC</span></div><span class="arr">→</span><div class="model">Carbon-3B</div><span class="arr">→</span><div class="io"><span class="tk dna">GCTAGC</span><span class="tk dna">TCGATC</span></div></div> | |
| <div class="row"><div class="io"><span class="tk txt">What</span><span class="tk txt">is</span><span class="tk txt">DNA</span><span class="tk txt">?</span></div><span class="arr">→</span><div class="model">Qwen3-14B</div><span class="arr">→</span><div class="io"><span class="tk txt">A</span><span class="tk txt">molecule</span><span class="tk txt">that</span><span class="tk txt">…</span></div></div> | |
| <div class="row cx"><div class="io"><span class="tk txt">What</span><span class="tk txt">does</span><span class="tk dna">TATAAA</span><span class="tk dna">GCTAGC</span><span class="tk txt">do</span><span class="tk txt">?</span></div><span class="arr">→</span><div class="model">Carbon-X</div><span class="arr">→</span><div class="io"><span class="tk txt">A</span><span class="tk txt">promoter:</span><span class="tk dna">TATAAA</span><span class="tk txt">…</span></div></div> | |
| </div> | |
| <figcaption><b>Figure 2.</b> The evolution from pure DNA models like Carbon-3B and pure text models such as Qwen3-14B to the hybrid Carbon-X model.</figcaption> | |
| </figure> | |
| <p>Earlier work such as <a class="inline" href="https://arxiv.org/abs/2505.23579" target="_blank" rel="noopener noreferrer">BioReason</a> connected the two by pairing a separate DNA encoder with a language-model decoder. Carbon-X instead trains a single 14B decoder on both DNA and English: DNA is simply another set of tokens in the same model.</p> | |
| <figure> | |
| <div class="arch-wrap"><div class="arch" role="img" aria-label="Left to right. The input 'Describe this sequence: ACGTACGATTCA' is split by a hybrid tokenizer into Qwen3 BPE text tokens and DNA 6-mer tokens wrapped in dna tags. Each token gets a vector from one shared embedding matrix: Qwen3's text embeddings plus new DNA embeddings. The vectors pass through the Qwen3-14B decoder layers and LM head, which output an interleaved sequence of text and DNA tokens."> | |
| <div class="panel"> | |
| <h4>Input</h4> | |
| <div class="field">Describe this sequence: <span class="seq">ACGTACGATTCA</span></div> | |
| </div> | |
| <span class="arr">→</span> | |
| <div class="panel"> | |
| <h4>Hybrid tokenizer</h4> | |
| <div class="groups"> | |
| <div class="group"><span class="vlab txt">Qwen3 BPE</span><div class="rows tokrows"><span class="tk txt">Describe</span><span class="tk txt">this</span><span class="tk txt">sequence</span><span class="tk txt">:</span></div></div> | |
| <div class="group"><span class="vlab dna">DNA 6-mers</span><div class="rows tokrows"><span class="tk dna"><dna></span><span class="tk dna">ACGTAC</span><span class="tk dna">GATTCA</span><span class="tk dna"></dna></span></div></div> | |
| </div> | |
| </div> | |
| <span class="arr">→</span> | |
| <div class="panel"> | |
| <h4>Shared embedding matrix</h4> | |
| <div class="groups"> | |
| <div class="group"><span class="vlab txt">Text (Qwen3)</span><div class="rows"><span class="vec txt"><i style="--s:48%"></i><i style="--s:53%"></i><i style="--s:25%"></i><i style="--s:33%"></i><i style="--s:45%"></i><i style="--s:49%"></i><i style="--s:23%"></i><i style="--s:72%"></i><i style="--s:55%"></i><i style="--s:43%"></i></span><span class="vec txt"><i style="--s:36%"></i><i style="--s:66%"></i><i style="--s:59%"></i><i style="--s:23%"></i><i style="--s:49%"></i><i style="--s:33%"></i><i style="--s:60%"></i><i style="--s:28%"></i><i style="--s:53%"></i><i style="--s:48%"></i></span><span class="vec txt"><i style="--s:29%"></i><i style="--s:21%"></i><i style="--s:76%"></i><i style="--s:78%"></i><i style="--s:25%"></i><i style="--s:46%"></i><i style="--s:77%"></i><i style="--s:30%"></i><i style="--s:44%"></i><i style="--s:52%"></i></span><span class="vec txt"><i style="--s:60%"></i><i style="--s:35%"></i><i style="--s:29%"></i><i style="--s:79%"></i><i style="--s:46%"></i><i style="--s:32%"></i><i style="--s:52%"></i><i style="--s:38%"></i><i style="--s:28%"></i><i style="--s:66%"></i></span></div></div> | |
| <div class="group"><span class="vlab dna">DNA (new)</span><div class="rows"><span class="vec dna"><i style="--s:40%"></i><i style="--s:40%"></i><i style="--s:23%"></i><i style="--s:75%"></i><i style="--s:33%"></i><i style="--s:73%"></i><i style="--s:47%"></i><i style="--s:33%"></i><i style="--s:55%"></i><i style="--s:43%"></i></span><span class="vec dna"><i style="--s:65%"></i><i style="--s:58%"></i><i style="--s:26%"></i><i style="--s:32%"></i><i style="--s:46%"></i><i style="--s:24%"></i><i style="--s:21%"></i><i style="--s:44%"></i><i style="--s:60%"></i><i style="--s:58%"></i></span><span class="vec dna"><i style="--s:77%"></i><i style="--s:38%"></i><i style="--s:45%"></i><i style="--s:49%"></i><i style="--s:45%"></i><i style="--s:70%"></i><i style="--s:46%"></i><i style="--s:45%"></i><i style="--s:34%"></i><i style="--s:33%"></i></span><span class="vec dna"><i style="--s:64%"></i><i style="--s:26%"></i><i style="--s:35%"></i><i style="--s:66%"></i><i style="--s:71%"></i><i style="--s:78%"></i><i style="--s:75%"></i><i style="--s:71%"></i><i style="--s:64%"></i><i style="--s:58%"></i></span></div></div> | |
| </div> | |
| </div> | |
| <span class="arr">→</span> | |
| <div class="panel core"> | |
| <h4>Qwen3-14B backbone</h4> | |
| <div class="layers"><div>Decoder layer 1</div><div>Decoder layer 2</div><div class="gap">⋯</div><div>Decoder layer N</div><div class="head">LM head</div></div> | |
| </div> | |
| <span class="arr">→</span> | |
| <div class="panel"> | |
| <h4>Output</h4> | |
| <div class="rows"><span class="tk txt">This</span><span class="tk txt">sequence</span><span class="tk txt">contains</span><span class="tk txt">a</span><span class="tk txt">promoter</span><span class="tk dna"><dna></span><span class="tk dna">TATAAA</span><span class="tk dna"></dna></span></div> | |
| </div> | |
| </div></div> | |
| <figcaption><b>Figure 3.</b> One model for DNA and English: a hybrid tokenizer and a shared embedding matrix feed the Qwen3-14B backbone.</figcaption> | |
| </figure> | |
| <p><strong>Application: predicting the effect of a mutation.</strong> Given an 8,192-base window of human DNA with a single A→G change, Carbon-X correctly identifies the variant as pathogenic and explains why: the change disrupts the AG at a splice site. Of the other models, only GPT-6.1 Sol at maximum effort got it right; Opus 5.5 reached the right label for the wrong reason, and the rest answered benign, were blocked, or ran out of tokens.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/vep-comparison.html" title="One variant, every model: NF1 c.6007-2A>G" loading="lazy"></iframe> | |
| <figcaption><b>Figure 4.</b> One variant, every model: NF1 c.6007-2A>G.</figcaption> | |
| </figure> | |
| </section> | |
| <section id="recipe"> | |
| <h2>How we built Carbon-X</h2> | |
| <p>We built Carbon-X by adapting <a class="inline" href="https://huggingface.co/Qwen/Qwen3-14B" target="_blank" rel="noopener noreferrer">Qwen3-14B</a> to DNA through midtraining, supervised fine-tuning, and reinforcement learning. Three challenges shaped the recipe: learning from genomes without losing language ability, building reliable supervision where existing models were weak teachers, and turning sequence knowledge into better answers using verifiable rewards.</p> | |
| <h3>Learning DNA without losing language</h3> | |
| <p><strong>Teaching Qwen3 DNA required balancing new capabilities against the language skills it already had.</strong> We started with a hybrid tokenizer, extending Qwen3’s vocabulary with 4,096 six-base DNA tokens, plus tokens for sequence boundaries and partial sequences. This gives genomic sequences a consistent representation instead of the irregular fragments produced by text tokenization. Six-base tokens also encode DNA more compactly than BPE tokenizers, using roughly a third as many tokens as Qwen3’s tokenizer.</p> | |
| <p>The design builds on our earlier <a class="inline" href="https://www.biorxiv.org/content/10.64898/2026.05.22.727119v1" target="_blank" rel="noopener noreferrer">Carbon-3B</a> results, where six-base tokenization outperformed byte-pair encoding, and follows <a class="inline" href="https://arxiv.org/abs/2502.07527" target="_blank" rel="noopener noreferrer">NatureLM</a>’s approach to combining biological sequences with text. Carbon-X processes both token types within a single model and can generate responses containing DNA and natural language.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/tokenizer.html" title="Qwen3 tokenizer vs Carbon-X tokenizer" loading="lazy"></iframe> | |
| <figcaption><b>Figure 5.</b> Qwen3 tokenizer vs Carbon-X tokenizer.</figcaption> | |
| </figure> | |
| <p><strong>A larger model, a lower learning rate, and sustained training gave us the best balance between DNA and language performance.</strong> Our 4B experiments showed that higher learning rates accelerated DNA learning but caused substantial forgetting on English tasks. We scaled the recipe to Qwen3-14B and trained on 300 billion tokens combining genomic sequences, biology text, and general English.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/progression-300b.html" title="Carbon-X midtraining on 300B tokens" loading="lazy"></iframe> | |
| <figcaption><b>Figure 6.</b> Carbon-X midtraining on 300B tokens.</figcaption> | |
| </figure> | |
| <details class="box"> | |
| <summary>Training data</summary> | |
| <div class="dbody"> | |
| <p>The final run used a 300B-token mixture made mostly of raw DNA from Carbon-3B’s data mixture, built from <a class="inline" href="https://huggingface.co/datasets/arcinstitute/opengenome2" target="_blank" rel="noopener noreferrer">OpenGenome2</a> and <a class="inline" href="https://huggingface.co/datasets/GenerTeam/pretrain_data_eukaryote" target="_blank" rel="noopener noreferrer">GENERator</a>, English text from web datasets and PDFs, and instruction data to help retain conversational and instruction-following abilities. The rest paired DNA with text: documents mined and filtered from web pages, papers and forums, plus synthetic examples in which Gemini analyzes annotated RefSeq sequences.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/data-mixture.html" title="Midtraining data mixture" loading="lazy"></iframe> | |
| <figcaption><b>Figure 7.</b> Midtraining data mixture.</figcaption> | |
| </figure> | |
| </div> | |
| </details> | |
| <details class="box"> | |
| <summary>Midtraining ablations</summary> | |
| <div class="dbody"> | |
| <p><strong>DNA and English favored different learning rates.</strong> In our midtraining ablations, higher learning rates improved DNA performance but caused substantial forgetting, while lower ones preserved English but slowed progress on DNA; no setting worked well for both. Training the randomly initialized DNA embeddings first, with the rest of the model frozen, helped a little but did not resolve the conflict.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/lr-sweep-4b.html" title="Qwen3-4B midtraining learning-rate sweep" loading="lazy"></iframe> | |
| <figcaption><b>Figure 8.</b> Qwen3-4B midtraining learning-rate sweep.</figcaption> | |
| </figure> | |
| <p>We tried different techniques for initializing DNA embeddings, separate learning rates for the embeddings and transformer body, protecting layers with conflicting DNA and English gradients, merging DNA and English specialists, and increasing English replay. None recovered strong performance in both domains.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/cpt-fixes.html" title="Continual-pretraining fixes we tried" loading="lazy"></iframe> | |
| <figcaption><b>Figure 9.</b> Continual-pretraining fixes we tried.</figcaption> | |
| </figure> | |
| <p><strong>What worked: a lower learning rate, a larger model and a longer run.</strong> Instead of the learning rate at which the model learned DNA fastest, we chose one at which DNA improved steadily, if more slowly, without losing English, and trained long enough to reach good performance on DNA tasks. At 4B, even these lower learning rates reduced English performance; at 14B, the same approach preserved most of it. We also used cosine decay: warmup–stable–decay keeps the learning rate high for most of training, which worsened forgetting in our experiments.</p> | |
| </div> | |
| </details> | |
| <p><strong>Midtraining unlocked gains that RL alone did not deliver.</strong> On FrontierVEP, RL applied directly to Qwen3-14B failed to rise above the 50% random-chance baseline throughout our training runs. Starting from the midtrained checkpoint, the same approach moved beyond chance and continued improving.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/frontiervep-base-mid.html" title="FrontierVEP: Base + RL vs Midtraining + RL" loading="lazy"></iframe> | |
| <figcaption><b>Figure 10.</b> FrontierVEP: base + RL vs midtraining + RL.</figcaption> | |
| </figure> | |
| <h3>Building supervision we could trust</h3> | |
| <p><strong>Higher benchmark scores did not guarantee grounded answers.</strong> The model still invented database references and misidentified genes, motivating supervision that could be checked against the supplied DNA.</p> | |
| <p>Standard distillation offered no reliable shortcut: frontier models frequently refused requests or hallucinated, while open models struggled with the underlying tasks. We initially used RL rollouts from our midtrained model for self-distillation, but these responses carried unsupported claims into the supervision.</p> | |
| <p><strong>We worked with domain experts to build deterministic verifiers that check claims against the input sequence.</strong> These checks helped us construct grounded training responses. The resulting SFT model achieved higher overall FrontierVEP performance than our self-distilled model.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/sft-subgroups.html" title="FrontierVEP SFT: self-distilled vs grounded" loading="lazy"></iframe> | |
| <figcaption><b>Figure 11.</b> SFT performance on FrontierVEP overall and across eight subgroups: self-distilled vs grounded checkpoints over five epochs. Panel y-axis limits vary.</figcaption> | |
| </figure> | |
| <h3>Improving with verifiable rewards</h3> | |
| <p><strong>We then used reinforcement learning to improve performance on tasks with checkable outcomes.</strong> Starting from the SFT model, we trained on mutation-effect prediction, DNA question answering, sequence recovery, and sequence-position tasks, alongside instruction-following tasks. Task-specific verifiers supplied the rewards.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/training-grid-v26.html" title="Carbon-X RL training rewards and evaluations (stage 1)" loading="lazy"></iframe> | |
| <figcaption><b>Figure 12.</b> Carbon-X RL training rewards and evaluations (stage 1).</figcaption> | |
| </figure> | |
| <p><strong>RL improved performance across all five DNA benchmarks</strong>, with the largest gains on mutation-effect prediction: approximately 20 percentage points on BioReason Non-SNV and 10 points on BioReason Coding over the SFT checkpoint.</p> | |
| </section> | |
| <section> | |
| <h2>Putting Carbon-X to work</h2> | |
| <p>Carbon-X can serve as a DNA specialist within a larger agent or run sequence-analysis tasks directly. Its low inference cost and support for local deployment make both approaches accessible on researchers’ own infrastructure.</p> | |
| <h3>A DNA specialist for other models</h3> | |
| <p>Using Carbon-X as a sub-agent improved Kimi K3’s performance in our agent evaluation, raising its DNA intelligence index from 49.4% to 60.5%—an improvement of 11.1 percentage points across five benchmarks.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/kimi-carbon-index.html" title="Kimi K3 DNA intelligence index with Carbon-X" loading="lazy"></iframe> | |
| <figcaption><b>Figure 13.</b> Kimi K3 alone versus Kimi K3 with required consultation of Carbon-X as a sub-agent. Both setups had Python available and were evaluated on 100 questions per benchmark across five DNA benchmarks. The intelligence index is the unweighted mean of the five scores.</figcaption> | |
| </figure> | |
| <h3>DNA analysis at scale</h3> | |
| <p>Large-scale genomic studies can require millions of predictions. We estimate that running inference on 2.4 million genetic variants whose effects on health remain uncertain would cost about $300 with Carbon-X, compared with $32,000 to $4 million using the frontier AI models evaluated, depending on the model and reasoning setting.</p> | |
| <figure> | |
| <iframe class="fig" src="figures/clinvar-cost.html" title="Projected cost to process 2.4M ClinVar VUS" loading="lazy"></iframe> | |
| <figcaption><b>Figure 14.</b> Projected cost to process 2.4 million ClinVar variants of uncertain significance.</figcaption> | |
| </figure> | |
| <h3>Running Carbon-X locally</h3> | |
| <p>Carbon-X works with llama.cpp, so you can ask questions about DNA sequences locally, without sending them to an external model provider. For researchers working with sensitive or unpublished sequences, this keeps the data within their own infrastructure. The demo below shows Carbon-X running locally on a MacBook.</p> | |
| <figure> | |
| <video controls playsinline preload="metadata" aria-label="Carbon-X running locally"> | |
| <source src="assets/carbon-x-demo-16s.mp4" type="video/mp4"> | |
| <a href="assets/carbon-x-demo-16s.mp4">Download the Carbon-X demo video</a>. | |
| </video> | |
| </figure> | |
| </section> | |
| <section> | |
| <h2>The importance of being open</h2> | |
| <p>As frontier labs move into biology, researchers should not have to depend on a handful of closed APIs. They need strong models they can run on their own hardware and afford to use at scale, while keeping sensitive data private.</p> | |
| <p><strong>Routine requests can trigger refusals.</strong> Safeguards matter in biology, but many closed models also refuse to answer legitimate biology questions. This has a measurable impact on benchmark performance. For example, on FrontierVEP, which tests whether mutations are benign or harmful, Gemini 4 Argon’s accuracy drops from 58% to 38% as reasoning effort increases, alongside a rise in refusals. We see a similar pattern with Claude Opus 5.5 on NextBase, which tests a model’s ability to generate DNA sequences: its score drops from 17.7% at low reasoning effort to just 1.1% at maximum effort.</p> | |
| <p>Performance also depends on whether a model attempts the task. Several frontier models refused a large share of legitimate genomic requests. For example, on the NextBase task, the refusal rates for Opus 5.5 and Gemini 4 Argon increased significantly with reasoning effort, leading to a corresponding drop in performance.</p> | |
| <figure> | |
| <div class="legend"> | |
| <span><i style="background:#2b6cd4"></i>GPT-6.1 Sol</span><span><i style="background:#bf7039"></i>Opus 5.5</span><span><i style="background:#8460a5"></i>Gemini 4 Argon</span> | |
| <span class="sep"><b class="ring"></b>Low reasoning</span><span><b class="dot"></b>Max reasoning</span> | |
| </div> | |
| <div class="chart" id="fig-refusal"></div> | |
| <figcaption><b>Figure 15.</b> NextBase score against refusal rate at low and max reasoning effort. Refusals are provider-reported content filter stops.</figcaption> | |
| </figure> | |
| </section> | |
| <section> | |
| <h2>Limitations</h2> | |
| <p>Carbon-X is a starting point for research on connecting DNA and language. We hope the model and training recipe help others investigate how knowledge learned from sequences can support reliable answers in natural language. Carbon-X has not been clinically validated and further work is needed to make its sequence-based answers consistently accurate and grounded.</p> | |
| <ul class="clean"> | |
| <li><b>Exact sequence operations:</b> The model can miscount bases, confuse coordinates, or make incorrect edits, particularly when locating exact positions in long sequences. One direction is to use it as a specialist within a larger assistant, combining Carbon-X’s sequence capabilities with tools that perform exact operations.</li> | |
| <li><b>Missing context and hallucinations:</b> The model may invent a sequence’s identity or function when the prompt does not provide enough information. Include known annotations and relevant context where available, and verify its biological claims.</li> | |
| <li><b>Short DNA tasks and general biology:</b> Frontier models remain stronger on short DNA exercises and textbook biology questions.</li> | |
| <li><b>Long sequence generation:</b> Quality degrades when the model is asked to generate long DNA sequences.</li> | |
| </ul> | |
| </section> | |
| <section> | |
| <h2>Safety and dual use</h2> | |
| <p>We're releasing Carbon-X to make DNA research more accessible and support local use for sensitive data. Openness also enables independent scrutiny of the model’s limitations and potential dual-use risks.</p> | |
| <p>With this in mind, we excluded eukaryote-infecting viral genomes from our DNA training data and evaluated Carbon-X on sequence-editing tasks with potential dual-use relevance. We found limited editing reliability and no advantage over readily available deterministic tools on those tasks. However, further research is needed to fully assess potential misuse risks, including after fine-tuning or integration with other systems.</p> | |
| </section> | |
| <section id="acknowledgements"> | |
| <h2>Acknowledgements</h2> | |
| <p>Carbon-X is a joint collaboration between the research teams at Hugging Face, Zhongguancun Academy, and the Telethon Institute of Genetics and Medicine/University of Naples “Federico II”.</p> | |
| <div class="partner-logos" aria-label="Research partners"> | |
| <a href="https://huggingface.co" target="_blank" rel="noopener noreferrer"><img src="assets/partners/hugging-face.png" alt="Hugging Face" width="1926" height="512" loading="lazy"></a> | |
| <a href="https://www.bza.edu.cn/en/" target="_blank" rel="noopener noreferrer" aria-label="Zhongguancun Academy"><svg viewBox="160 770 1730 510" role="img" aria-label="Zhongguancun Academy"><image href="assets/partners/zhongguancun-academy.png" width="2048" height="2048"/></svg></a> | |
| <a href="https://www.tigem.it/" target="_blank" rel="noopener noreferrer" aria-label="Telethon Institute of Genetics and Medicine"><svg viewBox="575 460 900 530" role="img" aria-label="Telethon Institute of Genetics and Medicine"><image href="assets/partners/tigem.jpg" width="2048" height="1448"/></svg></a> | |
| <a href="https://www.dicmapi.unina.it/" target="_blank" rel="noopener noreferrer"><img src="assets/partners/university-naples.png" alt="University of Naples Federico II" width="960" height="312" loading="lazy"></a> | |
| </div> | |
| </section> | |
| </main> | |
| <footer><div class="col">Hugging Face · Carbon-X · © 2026</div></footer> | |
| <script> | |
| /* Figure 1: score vs cost, native SVG. The arrow from Qwen3-14B to Carbon-X draws itself while a | |
| "thinking" label cycles through what the training actually felt like. */ | |
| (function () { | |
| const box = document.getElementById("fig1"); | |
| if (!box) return; | |
| // Exact values from the leaderboard snapshot in HuggingFaceBio/carbon-x-blog-figures (overall-performance.html). | |
| // x = cost per problem relative to Carbon-X; y = DNA intelligence index (%). Each model at its best reasoning effort. | |
| const PTS = [ | |
| { id: "carbon", logo: "carbon", effort: "none", name: "Carbon-X", short: "Carbon-X", x: 1.0, y: 66.69219775988992, hl: true }, | |
| { id: "qwen", logo: "qwen", effort: "none", name: "Qwen3-14B", short: "Qwen3-14B", x: 4.223377088306846, y: 39.96410541285459 }, | |
| { id: "deepseek", logo: "deepseek", effort: "low", name: "DeepSeek V4.1 Flash", short: "DeepSeek", x: 180.9590206842872, y: 47.107917709556894 }, | |
| { id: "kimi", logo: "moonshot", effort: "low", name: "Kimi K3", short: "Kimi K3", x: 688.0718929132777, y: 50.39465358277906 }, | |
| { id: "glm", logo: "zai", effort: "low", name: "GLM-5.3", short: "GLM-5.3", x: 1027.9301057416274, y: 39.83992062340639 }, | |
| { id: "gpt", logo: "openai", effort: "max", name: "GPT-6.1 Sol", short: "GPT-6.1", x: 1782.9613878159735, y: 67.51291997339764 }, | |
| ].map(p => ({ ...p, pct: p.y.toFixed(0) + "%" })); | |
| // The climb happens in equal steps: the arrow waits at each waypoint while that stage "thinks", | |
| // then moves at constant speed to the next one, where the next phrase takes over. | |
| const PHRASES = ["Building evaluations…", "Filtering the web…", "Tokenizing in 6-mers…", "Fighting forgetting…", | |
| "Despair…", "One more YOLO run…", "RLing…"]; | |
| const THINK = 1150; | |
| const STEPS = PHRASES.map((w, i) => ({ w, think: THINK, to: (i + 1) / PHRASES.length })); | |
| const MOVE = 650, HOLD = 3800; | |
| const GLYPHS = ["·", "✢", "✳", "✶", "✻", "✽", "✻", "✶", "✳", "✢"]; | |
| const SCHED = []; { let t = 0, from = 0; for (const s of STEPS) { SCHED.push({ ...s, from, t0: t, t1: t + s.think, t2: t + s.think + MOVE }); t += s.think + MOVE; from = s.to; } } | |
| const DRAW = SCHED[SCHED.length - 1].t2; | |
| const reduce = matchMedia("(prefers-reduced-motion: reduce)").matches; | |
| const ns = "http://www.w3.org/2000/svg"; | |
| let anim = null, t0 = 0, raf = 0; | |
| function render() { | |
| const W = box.clientWidth, narrow = W < 560; | |
| const H = Math.round(narrow ? W * 0.95 : Math.min(W * 0.6, 560)); | |
| const m = narrow ? { l: 50, r: 24, t: 30, b: 44 } : { l: 70, r: 24, t: 30, b: 54 }; | |
| const r = narrow ? 11 : 16, fsN = narrow ? 11.5 : 14, fsP = narrow ? 12.5 : 16, fsA = narrow ? 10.5 : 12.5; | |
| const LO = -0.32, HI = 4.08; // log10 domain, padded around 1x and 10,000x | |
| const X = v => m.l + (W - m.l - m.r) * (Math.log10(v) - LO) / (HI - LO); | |
| const Y = v => m.t + (H - m.t - m.b) * (75 - v) / 40; | |
| let s = ""; | |
| for (let v = 35; v <= 75; v += 5) { | |
| s += `<line x1="${m.l}" x2="${W - m.r}" y1="${Y(v)}" y2="${Y(v)}" stroke="var(--rule)"/>`; | |
| s += `<text x="${m.l - 8}" y="${Y(v) + 4}" text-anchor="end" font-size="${fsA}" fill="var(--ink-3)">${v}%</text>`; | |
| } | |
| ["1×", "10×", "100×", "1,000×", "10,000×"].forEach((t, i) => { | |
| const x = X(10 ** i); | |
| s += `<text x="${x}" y="${H - m.b + 20}" text-anchor="middle" font-size="${fsA}" fill="var(--ink-3)">${t}</text>`; | |
| }); | |
| s += `<text x="${(m.l + W - m.r) / 2}" y="${H - 6}" text-anchor="middle" font-size="${fsA + 0.5}" fill="var(--ink-2)">Cost per problem relative to Carbon-X</text>`; | |
| s += `<text transform="translate(${narrow ? 10 : 18} ${(m.t + H - m.b) / 2}) rotate(-90)" text-anchor="middle" font-size="${fsA + 0.5}" fill="var(--ink-2)">DNA intelligence index (%)</text>`; | |
| // arrow (animated parts are filled in by the frame loop) | |
| const q = PTS[1], c = PTS[0]; | |
| const ax = X(q.x), ay = Y(q.y), bx = X(c.x), by = Y(c.y); | |
| const len = Math.hypot(bx - ax, by - ay), ux = (bx - ax) / len, uy = (by - ay) / len; | |
| const A = { x: ax + ux * (r + fsP + 18), y: ay + uy * (r + fsP + 18) }, B = { x: bx - ux * (r + 8), y: by - uy * (r + 8) }; | |
| // waypoints zig-zag a little either side of the straight line; first and last sit on it | |
| const amp = narrow ? 4 : 7, n = PHRASES.length; | |
| const WP = Array.from({ length: n + 1 }, (_, i) => { | |
| const off = i === 0 || i === n ? 0 : amp * (i % 2 ? 1 : -1); | |
| return { x: A.x + (B.x - A.x) * i / n - uy * off, y: A.y + (B.y - A.y) * i / n + ux * off }; | |
| }); | |
| s += `<polyline id="f1-shaft" points="${WP.map(w => `${w.x},${w.y}`).join(" ")}" fill="none" stroke="var(--green)" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"/>`; | |
| s += `<polygon id="f1-head" points="0,0 -11,-5.5 -11,5.5" fill="var(--green)"/>`; | |
| // points | |
| PTS.forEach(p => { | |
| const x = X(p.x), y = Y(p.y), name = narrow ? p.short : p.name; | |
| const w = name.length * fsN * 0.56, right = x + r + 8 + w <= W - m.r; | |
| s += `<g class="pt" tabindex="0" aria-label="${p.name}, ${p.effort} reasoning, ${p.pct}, ${p.x.toFixed(2)} times Carbon-X cost"><title>${p.name} · ${p.effort} reasoning\n${p.pct}\n${p.x.toFixed(2)}× Carbon-X cost</title>`; | |
| s += `<image href="assets/logos/${p.logo}.png" x="${x - r}" y="${y - r}" width="${2 * r}" height="${2 * r}" preserveAspectRatio="xMidYMid meet" class="logo"/>`; | |
| s += `<text x="${x}" y="${y - r - 7}" text-anchor="middle" font-size="${fsP}" font-weight="700" fill="var(--ink)">${p.pct}</text>`; | |
| s += `<text x="${right ? x + r + 8 : x - r - 8}" y="${y + fsN * 0.36}" text-anchor="${right ? "start" : "end"}" font-size="${fsN}" font-weight="${p.hl ? 700 : 400}" fill="${p.hl ? "var(--ink)" : "var(--ink-2)"}">${name}</text></g>`; | |
| }); | |
| // thinking label, rides next to the arrow tip | |
| s += `<g id="f1-think" opacity="0"><text id="f1-glyph" font-size="${fsN + 2}" fill="var(--green)">✻</text><text id="f1-word" x="${fsN + 6}" font-size="${fsN}" font-style="italic" fill="var(--ink-2)"></text></g>`; | |
| const svg = document.createElementNS(ns, "svg"); | |
| svg.setAttribute("viewBox", `0 0 ${W} ${H}`); | |
| svg.setAttribute("role", "img"); | |
| svg.setAttribute("aria-label", "DNA intelligence index against cost per problem relative to Carbon-X. Carbon-X 67% at 1x; Qwen3-14B 40%; DeepSeek V4.1 Flash 47%; Kimi K3 50%; GLM-5.3 40%; GPT-6.1 Sol 68% at about 1,800 times the cost."); | |
| svg.innerHTML = s; | |
| box.replaceChildren(svg); | |
| anim = { WP, shaft: svg.getElementById("f1-shaft"), head: svg.getElementById("f1-head"), | |
| narrow, fixed: { x: X(30), y: Y(74) }, think: svg.getElementById("f1-think"), glyph: svg.getElementById("f1-glyph"), word: svg.getElementById("f1-word"), fsN }; | |
| frame(reduce ? DRAW : performance.now() - t0); | |
| } | |
| // progress along the arrow and the stage being worked on at time `cyc` | |
| function at(cyc) { | |
| const s = SCHED.find(s => cyc < s.t2) || SCHED[SCHED.length - 1]; | |
| if (cyc >= DRAW) return { p: 1, step: null }; | |
| if (cyc < s.t1) return { p: s.from, step: s }; | |
| const k = (cyc - s.t1) / MOVE; // linear | |
| return { p: s.from + (s.to - s.from) * k, step: s }; | |
| } | |
| function frame(el) { | |
| if (!anim) return; | |
| const { WP, shaft, head, think, glyph, word, fsN, narrow, fixed } = anim; | |
| const cyc = el % (DRAW + HOLD), st = reduce ? { p: 1, step: null } : at(cyc), p = st.p; | |
| // walk the zig-zag: completed segments plus the part of the current one | |
| // at a waypoint, stay on the segment just travelled so the head keeps pointing the way it came | |
| const n = WP.length - 1, seg = Math.max(0, Math.min(n - 1, Math.ceil(p * n - 1e-9) - 1)), k = Math.min(1, p * n - seg); | |
| const a = WP[seg], b = WP[seg + 1], tx = a.x + (b.x - a.x) * k, ty = a.y + (b.y - a.y) * k; | |
| shaft.setAttribute("points", [...WP.slice(0, seg + 1), { x: tx, y: ty }].map(w => `${w.x},${w.y}`).join(" ")); | |
| const dl = Math.hypot(b.x - a.x, b.y - a.y), ux = (b.x - a.x) / dl, uy = (b.y - a.y) / dl; | |
| const ang = Math.atan2(uy, ux) * 180 / Math.PI; | |
| head.setAttribute("transform", `translate(${tx + ux * 2} ${ty + uy * 2}) rotate(${ang})`); | |
| head.setAttribute("opacity", p > 0.02 ? 1 : 0); | |
| const thinking = !!st.step; | |
| think.setAttribute("opacity", thinking ? 1 : 0); | |
| if (thinking) { | |
| word.textContent = st.step.w; | |
| glyph.textContent = GLYPHS[Math.floor(el / 190) % GLYPHS.length]; | |
| // slow breathing: scale and opacity follow a sine, so the glyph swells rather than ticks | |
| const br = (Math.sin(el / 520) + 1) / 2, c = (fsN + 2) * 0.36; | |
| glyph.setAttribute("transform", `translate(${c} ${-c}) scale(${0.82 + 0.26 * br}) translate(${-c} ${c})`); | |
| glyph.setAttribute("opacity", 0.55 + 0.45 * br); | |
| think.setAttribute("transform", narrow ? `translate(${fixed.x - 60} ${fixed.y + 4})` : `translate(${tx + 16} ${ty + fsN * 0.4})`); | |
| } | |
| } | |
| function loop(now) { frame(now - t0); raf = requestAnimationFrame(loop); } | |
| render(); | |
| if (!reduce) { | |
| // start when the figure scrolls into view; pause when it leaves | |
| new IntersectionObserver(([e]) => { | |
| cancelAnimationFrame(raf); | |
| if (e.isIntersecting) { t0 = performance.now(); raf = requestAnimationFrame(loop); } | |
| }, { threshold: 0.35 }).observe(box); | |
| } | |
| let rw = box.clientWidth; | |
| addEventListener("resize", () => { if (box.clientWidth !== rw) { rw = box.clientWidth; render(); } }); | |
| })(); | |
| </script> | |
| <script> | |
| /* NextBase score vs refusal rate, native SVG. Data from carbon-x-blog-figures/score-vs-refusal.html: | |
| seqrecqa_varlen only; refusals are provider-reported content_filter stop reasons, denominator | |
| excludes technical errors. Snapshot 2026-10-21T13:31:46Z. */ | |
| (function () { | |
| const box = document.getElementById("fig-refusal"); | |
| if (!box) return; | |
| const DATA = [ | |
| { model: "GPT-6.1 Sol", effort: "low", color: "#2b6cd4", score: 31.597395146753232, refusal: 0.0, refused: 0, eligible: 600 }, | |
| { model: "GPT-6.1 Sol", effort: "max", color: "#2b6cd4", score: 32.88423559737861, refusal: 0.0, refused: 0, eligible: 600 }, | |
| { model: "Opus 5.5", effort: "low", color: "#bf7039", score: 17.71717851566719, refusal: 7.666666666666667, refused: 46, eligible: 600 }, | |
| { model: "Opus 5.5", effort: "max", color: "#bf7039", score: 1.1115857127488284, refusal: 70.33333333333333, refused: 422, eligible: 600 }, | |
| { model: "Gemini 4 Argon", effort: "low", color: "#8460a5", score: 31.204647685966314, refusal: 12.333333333333334, refused: 74, eligible: 600 }, | |
| { model: "Gemini 4 Argon", effort: "high", color: "#8460a5", score: 7.970855853822358, refusal: 82.16666666666667, refused: 493, eligible: 600 }, | |
| ]; | |
| function render() { | |
| const W = box.clientWidth, narrow = W < 560; | |
| const H = Math.round(narrow ? W * 0.9 : Math.min(W * 0.55, 520)); | |
| const m = narrow ? { l: 46, r: 30, t: 14, b: 44 } : { l: 70, r: 40, t: 16, b: 54 }; | |
| const fsA = narrow ? 10.5 : 12.5, fsL = narrow ? 12 : 14.5, rr = narrow ? 5.5 : 7; | |
| const X = v => m.l + (W - m.l - m.r) * v / 100, Y = v => m.t + (H - m.t - m.b) * (40 - v) / 40; | |
| let s = ""; | |
| for (let v = 0; v <= 40; v += 5) { | |
| s += `<line x1="${m.l}" x2="${W - m.r}" y1="${Y(v)}" y2="${Y(v)}" stroke="var(--rule)"/>`; | |
| s += `<text x="${m.l - 8}" y="${Y(v) + 4}" text-anchor="end" font-size="${fsA}" fill="var(--ink-3)">${v}%</text>`; | |
| } | |
| for (let v = 0; v <= 100; v += 20) s += `<text x="${X(v)}" y="${H - m.b + 20}" text-anchor="middle" font-size="${fsA}" fill="var(--ink-3)">${v}%</text>`; | |
| s += `<text x="${(m.l + W - m.r) / 2}" y="${H - 6}" text-anchor="middle" font-size="${fsA + 0.5}" fill="var(--ink-2)">Refusal rate (%)</text>`; | |
| s += `<text transform="translate(${narrow ? 10 : 18} ${(m.t + H - m.b) / 2}) rotate(-90)" text-anchor="middle" font-size="${fsA + 0.5}" fill="var(--ink-2)">NextBase score (%)</text>`; | |
| [...new Set(DATA.map(d => d.model))].forEach(name => { | |
| const rows = DATA.filter(d => d.model === name); | |
| s += `<polyline points="${rows.map(d => `${X(d.refusal)},${Y(d.score)}`).join(" ")}" fill="none" stroke="${rows[0].color}" stroke-width="2" opacity=".6"/>`; | |
| }); | |
| DATA.forEach(d => { | |
| const px = X(d.refusal), py = Y(d.score), low = d.effort === "low"; | |
| const dy = d.model === "GPT-6.1 Sol" && low ? fsL + 6 : -10; | |
| s += `<g class="pt" tabindex="0" aria-label="${d.model}, ${d.effort} reasoning: ${d.score.toFixed(2)}% NextBase score, ${d.refusal.toFixed(2)}% refusal rate">`; | |
| s += `<title>${d.model} · ${d.effort} reasoning\nScore: ${d.score.toFixed(2)}%\nRefusals: ${d.refusal.toFixed(2)}% (${d.refused} / ${d.eligible})</title>`; | |
| s += `<circle cx="${px}" cy="${py}" r="${rr}" fill="${low ? "var(--paper)" : d.color}" stroke="${d.color}" stroke-width="2.5"/>`; | |
| s += `<text x="${px + rr + 8}" y="${py + dy}" font-size="${fsL}" font-weight="600" fill="${d.color}">${d.score.toFixed(0)}%</text></g>`; | |
| }); | |
| const svg = document.createElementNS("http://www.w3.org/2000/svg", "svg"); | |
| svg.setAttribute("viewBox", `0 0 ${W} ${H}`); | |
| svg.setAttribute("role", "img"); | |
| svg.setAttribute("aria-label", "NextBase score against refusal rate at low and high reasoning effort. GPT-6.1 Sol has no refusals and scores 32% and 33%. Opus 5.5 falls from 18% at 8% refusals to 1% at 70% refusals. Gemini 4 Argon falls from 31% at 12% refusals to 8% at 82% refusals."); | |
| svg.innerHTML = s; | |
| box.replaceChildren(svg); | |
| } | |
| render(); | |
| let rw = box.clientWidth; | |
| addEventListener("resize", () => { if (box.clientWidth !== rw) { rw = box.clientWidth; render(); } }); | |
| })(); | |
| </script> | |
| <script> | |
| /* Figures from HuggingFaceBio/carbon-x-blog-figures are copied into figures/ and embedded same-origin: | |
| match the page background and size each frame to its content. */ | |
| (function () { | |
| const paper = () => getComputedStyle(document.documentElement).getPropertyValue("--paper").trim(); | |
| document.querySelectorAll("iframe.fig").forEach(f => { | |
| const fit = () => { | |
| try { | |
| const d = f.contentDocument; | |
| if (!d || !d.body) return; | |
| // height of the actual content: lowest bottom edge among the body's visible children | |
| let h = 0; | |
| for (const el of d.body.children) { | |
| if (/^(SCRIPT|STYLE|TEMPLATE)$/.test(el.tagName)) continue; | |
| const r = el.getBoundingClientRect(), mb = parseFloat(d.defaultView.getComputedStyle(el).marginBottom) || 0; | |
| h = Math.max(h, r.bottom + mb); | |
| } | |
| h = Math.ceil(h + 2); | |
| if (h > 0) f.style.height = h + "px"; | |
| } catch (e) { /* cross-origin preview: keep the default height */ } | |
| }; | |
| f.addEventListener("load", () => { | |
| try { | |
| const d = f.contentDocument, st = d.createElement("style"); | |
| st.textContent = `:root{--paper:${paper()}} html,body{background:transparent!important;height:auto!important;min-height:0!important;margin:0} | |
| html{overflow-y:hidden!important;scrollbar-gutter:auto!important} | |
| body > .wrap, body > .fig {padding-top:0!important;padding-bottom:0!important} | |
| body > .wrap > figure {margin-top:0!important;margin-bottom:0!important} | |
| body > .fig:has(> .replay) {padding-top:34px!important}`; | |
| d.head.appendChild(st); | |
| new ResizeObserver(fit).observe(d.body); | |
| } catch (e) {} | |
| fit(); setTimeout(fit, 800); setTimeout(fit, 2500); | |
| }); | |
| }); | |
| })(); | |
| </script> | |
| </body> | |
| </html> | |