Hugging Face has released Carbon-X, an open DNA language model with 14.8B parameters. Carbon-X is a decoder-only model from Hugging Face Bio. It reads raw DNA in a chat and answers in English. It predicts whether mutations are benign or pathogenic, flags regulatory regions, classifies organisms and predicts next bases. The weights ship under Apache 2.0, so labs can keep sensitive sequences on their own GPUs. Hugging Face built it with Zhongguancun Academy and TIGEM / University of Naples Federico II.
TL;DR
- Size: 14.8B parameters (dense, no MoE), 40,960-token context.
- Runs on: 1 x 80 GB GPU in bfloat16 via vLLM or Transformers. Official GGUF: Q8_0 is 15.8 GB, Q4_K_M is 9.6 GB.
- Performance: 66.7% average on 5 DNA benchmarks, best of every open model tested (self-reported).
- Best: 74.9% on BioReason Coding, ahead of every model tested, open or closed.
- Worst: 29.8% on GenomeQA splice-site questions.
What is Carbon-X and how does it work?
Carbon-X starts from Qwen3-14B. DNA and text share 1 token stream in 1 decoder. There is no separate DNA encoder, unlike earlier work such as BioReason.
The key structure piece is a hybrid tokenizer. It keeps Qwen3’s BPE vocabulary for text and adds 5,515 DNA tokens. These include 4,096 non-overlapping 6-mers, 1,364 tail tokens and 3 markers. One DNA token covers 6 base pairs. DNA runs of 6 or more A/C/G/T bases are tagged automatically.
The backbone is a standard Qwen3 stack. It has 40 layers, hidden size 5120 and GQA with 8 KV heads.
How was Carbon-X trained?
Training ran in four stages, per the model card:
- Embedding warmup: about 10B tokens with the transformer frozen, so only the new DNA embeddings learn.
- Mid-training: 302B tokens at 8,192 sequence length. The mix is 55% raw DNA, 20% English, 15% biology text and 10% instruction data.
- SFT: 3 epochs over 20 dataset subsets, with responses checked by deterministic verifiers built with domain experts.
- RL: asynchronous GRPO with verifiable rewards over 2 rounds, using TRL and vLLM.
As per the release post, learning rate was the hardest knob. High rates learned DNA but forgot English. A lower rate, a 14B model and a long run fixed the trade-off. RL applied directly to Qwen3-14B never beat the 50% chance baseline on FrontierVEP.
How does Carbon-X perform on DNA benchmarks?
All scores are self-reported by Hugging Face. Carbon-X ran with thinking disabled. Competitors ran at their best reasoning effort.
| Benchmark | Carbon-X | GPT-6.1 Sol (Max) | Opus 5.5 (Max) | Kimi K3 (Low) |
|---|---|---|---|---|
| BioReason Coding | 74.9 | 71.0 | 70.3 | 53.9 |
| BioReason Non-SNV | 85.1 | 91.7 | 94.8 | 65.3 |
| GenomeQA | 63.6 | 67.2 | 62.2 | 52.0 |
| NextBase | 39.4 | 32.9 | 1.1 | 30.6 |
| FrontierVEP | 70.4 | 74.9 | 76.2 | 50.1 |
| Average | 66.7 | 67.5 | 60.9 | 50.4 |
Carbon-X sits 27 points above Qwen3-14B and 0.8 points below GPT-6.1 Sol. It leads on coding-variant effects and next-base prediction. Frontier models still win on complex variants and single-base variants.
Refusals shape these numbers. On NextBase, Opus 5.5 falls from 17.7% at low effort to 1.1% at max effort as refusals rise. Gemini 4 Argon drops from 58% to 38% on FrontierVEP as effort increases.
Carbon-X also keeps general skills: 79.3% prompt-level strict accuracy on IFEval.
‘+s.p+’
‘;
if(s.mix){h+=’
‘+s.mix.map(function(m){return ‘‘}).join(”)+’
‘+s.mix.map(function(m){return ‘‘+m[2]+’‘}).join(”)+’
‘}
det.innerHTML=h;if(s.mix)setTimeout(function(){det.querySelectorAll(‘.mix i’).forEach(function(x){x.style.width=x.dataset.w+’%’})},40);setTimeout(post,60)}
sts.forEach(function(s){s.onclick=function(){showSt(+s.dataset.s)}});showSt(0);
// benchmarks
var M=[‘Carbon-X’,’GPT-6.1 Sol (Max)’,’Opus 5.5 (Max)’,’Gemini 4 Argon (Low)’,’Kimi K3 (Low)’,’DeepSeek V4.1 Flash (Low)’,’GLM-5.3 (Low)’];
var B={‘Average’:[66.7,67.5,60.9,49.8,50.4,47.1,39.8],’BioReason Coding’:[74.9,71.0,70.3,6.5,53.9,51.4,52.1],’BioReason Non-SNV’:[85.1,91.7,94.8,90.3,65.3,55.9,60.1],’GenomeQA’:[63.6,67.2,62.2,63.1,52.0,49.7,44.6],’NextBase’:[39.4,32.9,1.1,31.2,30.6,27.2,9.7],’FrontierVEP’:[70.4,74.9,76.2,58.0,50.1,51.3,32.7]};
var N={‘Average’:’Carbon-X trails GPT-6.1 Sol by 0.8 points and sits 27 points above Qwen3-14B.’,’BioReason Coding’:’Coding-variant effects, 1,018 variants. Carbon-X leads every model tested.’,’BioReason Non-SNV’:’Insertions, deletions and other non-SNV changes, 873 variants. Opus 5.5 leads.’,’GenomeQA’:’2,600 multiple-choice questions over 6 tasks. GPT-6.1 Sol leads.’,’NextBase’:’Next-base prediction, 600 sequences. Carbon-X leads; Opus 5.5 collapses as refusals rise.’,’FrontierVEP’:’Single-base variants with up to 8,192-base context, 1,000 variants. Opus 5.5 leads.’};
var cur=”Average”,bs=document.getElementById(‘cx-bsel’),bars=document.getElementById(‘cx-bars’);
Object.keys(B).forEach(function(k){var b=document.createElement(‘button’);b.textContent=k;if(k===cur)b.className=”on”;b.onclick=function(){bs.querySelectorAll(‘button’).forEach(function(x){x.className=””});b.className=”on”;cur=k;drawBars(k)};bs.appendChild(b)});
function drawBars(k){var v=B[k];bars.innerHTML=M.map(function(m,i){var me=i===0?’ me’:”;return ”}).join(”);
setTimeout(function(){bars.querySelectorAll(‘.fl’).forEach(function(f){f.style.width=f.dataset.w+’%’})},40);document.getElementById(‘cx-bnote’).textContent=N[k];setTimeout(post,60)}
drawBars(cur);
// cost
function anim(el,to,fmt){var t0=null;function f(ts){if(!t0)t0=ts;var p=Math.min(1,(ts-t0)/1000),e=1-Math.pow(1-p,3);el.textContent=fmt(to*e);if(p<1)requestAnimationFrame(f)}requestAnimationFrame(f)}
function countUp(){anim(document.getElementById(‘cx-c1’),300,function(x){return ‘$’+Math.round(x)});anim(document.getElementById(‘cx-c2’),1800,function(x){return ‘~’+Math.round(x).toLocaleString(‘en-US’)+’x’});anim(document.getElementById(‘cx-c3’),11.1,function(x){return ‘+’+x.toFixed(1)})}
// limits
R.querySelectorAll(‘.lim .box’).forEach(function(b){b.onclick=function(){b.classList.toggle(‘open’);setTimeout(post,380)}});
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
