Hugging Face released Carbon-X on October 22, 2026, an open model you can ask questions about DNA in plain English: paste in a sequence, ask whether a mutation looks harmless or disease-causing, and it answers in words. Why it matters: it puts a near-frontier genetics assistant on researchers’ own hardware for free, where established DNA models answer in numbers or stay behind cloud APIs.
October 22, 2026
What happened
Hugging Face built Carbon-X with researchers at Zhongguancun Academy in Beijing and the Telethon Institute of Genetics and Medicine and the University of Naples Federico II in Italy, and released the weights under the Apache 2.0 license, so anyone can download, modify and run it. The model starts from Alibaba’s open Qwen3-14B and was then trained on 302 billion tokens mixing raw DNA, English and biology text. Its tokenizer reads DNA in six-letter chunks alongside ordinary words, so DNA and English share a single conversation.
What it can do — and what the numbers actually say
Carbon-X takes in about 49,000 base pairs at a time — a slice of a genome, not a whole one. Within that window it predicts whether mutations are benign or pathogenic, identifies regions that control gene activity, classifies sequences by organism group and predicts the next bases. That chat format is the break from its rivals: Google DeepMind’s AlphaGenome predicts mutation effects as numbers, while Evo 2, from the Arc Institute, Nvidia and academic partners, generates whole genomes. On five DNA benchmarks, Carbon-X averaged 66.7 percent, just behind GPT-6.1 Sol at 67.5 percent and ahead of every other open model the team tested, at roughly 1/1,800 of Sol’s cost per problem by Hugging Face’s own measurements. Treat those figures as the team’s claims, not settled results: Hugging Face built two of the five benchmarks itself, and closed models lost points in part by refusing genetics questions.
Built to stay local — and to leave viruses out
Senior research scientist Lewis Tunstall told WIRED the team kept the model at 14.8 billion parameters so it can run on a lab’s own hardware — it fits on a single 80GB GPU — keeping sensitive sequences off cloud services entirely, with nothing sent to Hugging Face. The team also left viruses out of the training data on purpose, to cut the chance of dangerous uses. The intended work is practical: judging whether a mutation is worth targeting with a drug or genetic therapy, where picking the wrong one wastes a program or worse.
Why it matters
For researchers, Carbon-X is the first DNA model that behaves like the chatbots they already use: no specialist pipeline, no numeric output to decode, no patient sequence uploaded to someone else’s server — and it is free. The honest caveat is in Hugging Face’s own model card: Carbon-X has not been clinically validated or tested in a wet lab, it can miscount bases and guess when given too little context, and it should not be used for medical decisions. Everything AI, tested: the chat interface is here today; whether its answers survive contact with a real lab is the result that actually counts, and that test has not been run yet.
Frequently asked questions
What is Carbon-X?
An open 14.8-billion-parameter model from Hugging Face and partners in China and Italy that reads raw DNA and answers questions about it in plain English, released October 22, 2026 under Apache 2.0.
Can Carbon-X read a whole human genome?
No. It handles about 49,000 base pairs per query, so researchers work with small sections of a genome at a time.
Has Carbon-X been validated in a laboratory?
No. Hugging Face states it has not been clinically validated or wet-lab tested and should not be used for medical decisions; its benchmark scores are the team’s own measurements.
How is Carbon-X different from AlphaGenome?
AlphaGenome returns numerical predictions about sequences and mutations. Carbon-X works conversationally: you ask in English and it answers in English, and it is small enough to run locally on one 80GB GPU.
Sources: Hugging Face Carbon-X model card, October 22, 2026; WIRED interview with Lewis Tunstall, via Superpower Daily, RuntimeWire and Madrobot coverage of the launch, October 22, 2026.