Hugging Face's model lets you chat with DNA⁠.

Hugging Face, the platform best known as a library of ready made AI models, on Thursday released Carbon-X, an open-source model for understanding DNA¹. It is trained on genomic sequences and on ordinary language, so a researcher can ask about the effects of a change in DNA and get a conversational answer back¹. Lewis Tunstall, a senior research scientist at the company, calls it being able to "chat with DNA"². Business Insider also reports it as "first of its kind" in combining DNA analysis and generation with plain language².

The model was built with the University of Naples Federico II in Italy and Zhongguancun Academy in China¹. The team took an existing language model, Qwen3-14B, and trained it further on DNA from humans, other mammals, plants and fungi². Wired says the training ran to 300 billion tokens mixing DNA, English text and examples linking the two¹. Viruses were left out of the training data on purpose, to avoid misuse².

The example use is a mutation. A researcher supplies a DNA sequence and asks whether a particular mutation could raise a person's risk of disease, and the model gives a prediction and an explanation². It can also generate new DNA sequences, which could help in developing gene therapies². It cannot read a whole genome: it takes sequences of roughly 50,000 base pairs, so users query small parts at a time¹.

The detail that matters most to a business is where it runs. At 14 billion parameters it is small enough to run on the user's own hardware, and no data is sent to Hugging Face or any other company¹. Tunstall says the team designed it that way because researchers face restrictions on uploading sensitive patient data to frontier AI models¹. He also says the point was to give open researchers tools that do not leave them "fully dependent on some external single provider"².

There is an awkward line in the small print. The developers say they have not yet validated the model with wet lab experiments, so they cannot say how accurate its predictions are¹. Nvidia agreed in September to buy Hugging Face for nearly $13 billion².

Small, local and not yet proven

Most of our readers do not sequence genomes, and nothing in the reporting mentions a British lab, a British regulator or a price in pounds. The reason to notice it is the shape of the thing. A capable specialist model that sits on your own machine and never sends data out is the answer to the question every finance or operations director asks first about AI: where does our data go. The same pattern, a small model trained hard on one subject and run in house, is what a firm with sensitive contracts, payroll or customer records should be asking its suppliers for.

Hold the excitement until the lab work is done. A model that explains a mutation fluently and wrongly is worse than no model, and its makers say plainly they do not yet know how often it is right. A confident answer is not a checked one, in biology or in your accounts.

Also today

• 1Password lets AI agents sign in without seeing passwords

  1Password for Agents lets an AI agent sign into your accounts without handling the password itself, and you still have to approve each sign-in request³.

• DeepSeek reportedly raises $15bn days after US restrictions

  The Chinese AI lab has reportedly raised about $15 billion, two days after the US ordered federal civilian agencies to remove its technology over security concerns⁴.

• Google ordered to stop work on Finnish datacentre sites

  Google has been ordered to stop work at two sites in Finland after an estimated quarter of a million trees were felled across 330 hectares for datacentres⁵.

• OpenAI agents analyse a mouse brain atlas in 32 minutes

  Thousands of OpenAI agents worked through a whole-brain CRISPR atlas of mouse cells in 32 minutes, work the team estimates would take months to a year by hand⁶.

• Palantir says NATO's Maven system has no kill switch

  Palantir's head of UK and European operations says neither the company nor the US government can remotely shut down Maven, the battlefield AI used by NATO's top command⁷.