Hugging Face's open Carbon-X model answers questions about DNA, nearly matching GPT-6.1 Sol

A 14B open model comes within a point of GPT-6.1 Sol on DNA tests, and Hugging Face says the closed models kept refusing.

By Think Facility · October 22, 2026 · Hugging Face Open weights

On October 22, 2026, Hugging Face released Carbon-X, an open model you can paste raw DNA into and question in plain English. It's built with Zhongguancun Academy in Beijing and the Telethon Institute of Genetics and Medicine at the University of Naples Federico II, and the weights are on Hugging Face under Apache 2.0.

Ask it whether a mutation is likely harmful and it answers, then explains why. Across five DNA benchmarks it averages 66.7%, against 67.5% for GPT-6.1 Sol, at what Hugging Face puts at about 1/1,800 of the cost per problem.

Size: 14.8B parameters, starting from Qwen3-14B
Training: 302B tokens of DNA, biology text, English and instruction data
DNA benchmark average: 66.7%, against 67.5% for GPT-6.1 Sol
Hardware: a single 80 GB GPU
License: Apache 2.0

One model for DNA and English

Earlier DNA models, including Hugging Face's own Carbon-3B, could score a sequence or predict the next bases, but they couldn't explain a prediction or answer a question. Chat models can talk biology but mostly learned it from descriptions. Carbon-X puts both in one decoder. DNA is read in six-base chunks, each its own token, alongside ordinary text.

A man in a blue T-shirt with a white earbud, sitting in a white mesh office chair, speaking to the camera in a video call — Hugging Face CEO Clem Delangue, interviewed on SiliconANGLE's theCUBE. Photo: SiliconANGLE theCUBE, CC BY 3.0, via Wikimedia Commons

Getting there was the hard part, by the team's account. In their tests, a high learning rate taught the model DNA but made it forget English, and a low one kept English but left DNA at chance. What worked was a larger model, a lower rate and a longer run. They also found that reinforcement learning applied straight to Qwen3-14B never got above chance on their variant test, FrontierVEP, until they did the DNA training first.

How it compares

Carbon-X leads every model Hugging Face tested on coding-variant effects and next-base prediction. GPT-6.1 Sol and Opus 5.5 beat it on complex variants and single-base variant effects. On next-base prediction, Opus 5.5 at maximum effort scored 1.1%.

Hugging Face blames refusals for that. Several frontier models refused a large share of legitimate genomic requests, the team says, and refused more as reasoning effort went up. Gemini 4 Argon's accuracy on FrontierVEP fell from 58% to 38% as effort rose.

As frontier labs move into biology, researchers should not have to depend on a handful of closed APIs.

Carbon-X: When Natural Language Meets DNA

Used as a helper, it also lifted another model. With Carbon-X as a required sub-agent, Kimi K3's score across the five DNA benchmarks went from 49.4% to 60.5%.

What it could mean for genetic test results

The number that reaches patients is the cost one. Hugging Face estimates that running Carbon-X over the 2.4 million variants of uncertain significance in ClinVar would cost about $300, against $32,000 to $4 million with the frontier models it tested.

Those are the genetic changes a test finds without enough data to say whether they matter. The National Cancer Institute says a VUS result is typically not considered in making health care decisions, and that most get reclassified as benign later. The National Human Genome Research Institute adds that there's far more genomic data on people of European ancestry, which leaves more of these unresolved results for everyone else.

Carbon-X doesn't settle any of them. Hugging Face says it hasn't been clinically validated and shouldn't be used for medical decisions, and the launch article describes no lab testing of its calls. What it changes is who can afford to try. I'd watch for the first group that runs it over a hospital's own unresolved variants and publishes how often it agrees with later reclassifications.

Privacy is the other half. Because the weights fit on one 80 GB GPU, a lab can keep sensitive or unpublished sequences on its own machines. The launch article shows it running on a MacBook through llama.cpp, though the model card says llama.cpp can't load the DNA tokenizer, so check which version you're running.

Safety and limits

The team left genomes of viruses that infect eukaryotes out of the training data. They tested Carbon-X on sequence-editing tasks with possible dual-use value and found it no better than ordinary tools. They say more work is needed on misuse, including after fine-tuning. It can miscount bases, may invent what a sequence does when the prompt lacks context, and wasn't trained to design new DNA.