Hugging Face’s launch art for Carbon-X. Image: Hugging Face
Hugging Face has released Carbon-X, a free, open AI model that reads raw DNA and talks about it in plain English. Paste in genetic code, ask whether a mutation is harmless or disease-causing, and it answers like a chatbot, explaining its reasoning.
The 14.8 billion-parameter model is out under the Apache 2.0 licence, so anyone can download it, change it and run it on their own machines. Hugging Face built it with researchers at Zhongguancun Academy in Beijing and the Telethon Institute of Genetics and Medicine and the University of Naples Federico II in Italy.
‘Being able to chat with DNA’
Lewis Tunstall, a senior research scientist at Hugging Face, told Wired that the model is meant to help researchers design new drugs and genetic therapies, and could one day help patients make sense of their own genomes:
We’ve unlocked a new capability in language models, which is fundamentally about being able to chat with DNA.
Lewis Tunstall, senior research scientist, Hugging Face, to Wired
Because the questions are in ordinary language, he added, even nontechnical users “can actually ask questions about DNA that wouldn’t be accessible through any other tool.”
How it learned to read genes
Carbon-X starts from Qwen3-14B, an open model from Alibaba, and was then trained on 302 billion tokens, more than half of them raw DNA. Its tokenizer reads DNA in six-letter chunks, the team’s launch post says. The genomes came from mammals (humans included), other vertebrates, invertebrates, plants, fungi, protozoa and microbes, plus English and biology text.
It can only take in about 49,000 DNA letters at a time, so you can’t drop a whole human genome into it, only small parts.
Close to GPT-6.1 Sol, at a fraction of the cost
On five DNA tests, two of which Hugging Face built itself, Carbon-X averaged 66.7%. That is just behind OpenAI’s GPT-6.1 Sol on maximum reasoning (67.5%) and ahead of Claude Opus 5.5 (60.9%), Gemini 4 Argon (49.8%) and every open model tried. It led all of them at judging protein-coding mutations (74.9%) and at predicting the next letters of a sequence (39.4%).
Hugging Face puts the cost per problem at about 1/1,800 of GPT-6.1 Sol’s, and estimates that checking the 2.4 million “uncertain” variants in the US ClinVar database would cost about $300 with Carbon-X, against $32,000 to $4 million with the frontier models it tested. These are the company’s own figures and haven’t been independently checked.
Part of the gap comes from refusals. The launch post says closed models often turn down legitimate genetics questions: Opus 5.5’s next-base score falls from 17.7% at low reasoning effort to 1.1% at maximum, and Gemini 4 Argon’s mutation accuracy drops from 58% to 38% as it thinks harder and refuses more.
Viruses left out on purpose
The team removed viruses that infect animals, plants and other complex life from the DNA it trained on, Tunstall said, because it didn’t want the model used for dangerous work. Arc Institute took the same step with its Evo 2 DNA model, which Google DeepMind has since used to test watermarks for AI-designed proteins and DNA. Tests on DNA-editing tasks that could be misused found it unreliable and no better than ordinary software, though Hugging Face says more research is needed.
The model card is blunt about limits: Carbon-X has not been clinically validated or tested in a wet lab and should not be used for medical decisions. It can miscount bases, invent what a sequence does when given too little context, and wasn’t trained to design new DNA.
Runs on your own hardware
Tunstall said the team kept the model small so labs can keep sensitive patient data off cloud services; nothing is sent to Hugging Face or anyone else. In full precision it fits on one 80GB data-centre GPU, and the launch post shows it running on a MacBook. The weights are on the Hugging Face Hub, open to download worldwide, with a ready-made skill so Claude Code and Codex CLI can call it.
It joins a busy month for AI in biology, from a Claude-designed antivenom that two labs say works to OpenAI’s agents working through a mouse brain atlas. It is also one of Hugging Face’s last big releases as an independent company: Nvidia agreed to buy it for about $13 billion in September, a deal expected to close in the first half of 2027.
Why it matters
Carbon-X puts a near-frontier genetics assistant on researchers’ own machines for free, with no closed chatbot refusing their questions. Whether its answers hold up in the lab is the test that hasn’t been run yet.
Sources: Hugging Face (Carbon-X model card), Hugging Face (launch post), Wired (interview with Lewis Tunstall), Bloomberg Law.
Latest More Labs news
