Hugging Face releases Carbon-X, an open-source AI that answers DNA questions

Overview

Updated Nov 3

Hugging Face released Carbon-X on October 22, an open model that reads DNA and answers questions about it in English. At 14.8 billion parameters it fits on a single graphics card, so a lab can query patient sequences without sending them to a cloud service.

The first outside test, published October 29 by InferenceKey, ran the model on a 24 GB gaming card and got 75% of 40 ClinVar variants right, close to the 74.9% Hugging Face reports. It also caught 19 of 20 harmful variants but wrongly flagged 8 of 20 harmless ones, and its written reasoning sometimes named the wrong DNA letter. No laboratory experiment has checked a prediction, and Hugging Face says the model should not be used for medical decisions.

Why it matters

A lab can now run a DNA-reading model on a gaming card it already owns, keeping patient sequences inside the building.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

14.8B

Model parameters

Fine-tuned from Qwen3-14B. Independent testers ran it on a 24 GB RTX 4090 with 8-bit weights and on a 48 GB RTX A6000 at full precision.

66.7%

Hugging Face's average on five DNA benchmarks

Second to GPT-6.1 Sol at 67.5% and ahead of Opus 5.5 at 60.9%. The company puts inference at roughly 1/1,800 of frontier API cost per problem, against 1,000x in its own blog subtitle.

75%

Independent score on 40 ClinVar variants, 24 GB card

InferenceKey's result with randomness off, against 80% on a 48 GB card. On 40 variants the true figure could sit anywhere from about 64% to 90%.

$300

Hugging Face's estimate to analyze ClinVar variants

All 2.4 million variants of uncertain significance, against $32,000 to $4 million the company estimated for frontier models. Assumes $5 per GPU-hour.

29.8%

Score on splice-site tasks

The weakest of the GenomeQA subtasks in the developers' breakdown. Histone marks score 34.6%.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Lewis Tunstall

Senior research scientist at Hugging Face

Led the Carbon-X project

Diego di Bernardo

Vice director, Telethon Institute of Genetics and Medicine; professor, University of Naples Federico II

Conceived Carbon-X and proposed it to Hugging Face

Maurizio Fiusco

PhD student, University of Naples Federico II

Contributed to Carbon-X development and testing

Organizations Involved

Hugging Face

AI model hosting platform

Developer and publisher of Carbon-X

Open-source AI platform that developed and released Carbon-X.

Telethon Institute of Genetics and Medicine

Research institute

Research partner in Carbon-X collaboration

Italian research institute focused on rare genetic diseases that co-developed Carbon-X.

Zhongguancun Academy

Research academy

Research partner in Carbon-X collaboration

Beijing research academy that co-developed Carbon-X.

Timeline

2 events Latest: October 29th, 2026 · 1 week ago

October 2026
First independent test of Carbon-X published
Latest Research

InferenceKey runs Carbon-X on a 24 GB RTX 4090 and a 48 GB RTX A6000, scoring 75% and 77.5% on 40 ClinVar variants. Labels match Hugging Face's numbers; explanations often name the wrong DNA letter or the wrong gene.
October 29th, 2026
Hugging Face releases Carbon-X
Product Release

Hugging Face launches Carbon-X, an open model trained on 302 billion tokens of DNA and natural language that answers questions about genetic sequences in English. Its model card says it has not been clinically validated.
October 22nd, 2026

Scenarios

Wet-lab studies validate Carbon-X's mutation predictions

Possible Resolves by Oct 22, 2027

Discussed by: Hugging Face developers and Wired

Independent research groups test Carbon-X's predictions about whether mutations are benign or pathogenic. If the model matches experimental results, it becomes a screening tool for drug development and genetic therapy design. Tunstall says the tool was built for exactly this purpose.

Validation tests reveal accuracy gaps in Carbon-X

Possible Resolves by Oct 22, 2027

Discussed by: Model card's own warning that Carbon-X is not clinically validated or tested in wet labs

Wet-lab experiments show Carbon-X's predictions diverge from actual outcomes for certain mutation types or gene regulatory sequences. Developers have explicitly cautioned that predictions are untested, and the model card warns against medical use. If errors surface, Carbon-X remains an experimental research tool rather than a diagnostic aid.

Carbon-X reaches clinical or personalized-medicine use

Unlikely Resolves by Oct 22, 2028

Discussed by: Lewis Tunstall in Wired interview

If validation succeeds and regulators follow, patients could eventually use the model to interpret their own genomes. Tunstall calls personalized medicine 'the holy grail.' This requires clinical validation, regulatory clearance, and a gene-therapy pipeline that uses the model — the longest path of the three.

Outside labs reproduce or fail to reproduce Carbon-X's 66.7% average

Possible Resolves by Apr 30, 2027

Discussed by: Independent reviewers at mlllm.io, who note the developers wrote two of the five benchmarks themselves

Hugging Face has not published a technical report or released evaluation code with a contamination check, and two of the five benchmarks were built by the developers. If outside groups rerun the tests on held-out DNA and get similar scores, Carbon-X becomes a credible cheap alternative to API models for variant triage. If scores fall, the 66.7% figure stays a vendor number.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

July 2021

AlphaFold 2 open-source release (2021)

DeepMind's AlphaFold 2 predicted protein structures with near-experimental accuracy, and the company open-sourced both the code and more than 200 million predicted structures. Structural biologists gained instant access to results that once took months in a lab.

Then

Researchers worldwide began using AlphaFold predictions for drug discovery and disease research within weeks.

Now

Open-source AI became a standard tool in molecular biology, with AlphaFold cited in thousands of papers.

Why this matters now

Like AlphaFold, Carbon-X is an open AI tool aimed at a hard biological problem. But AlphaFold required Google-scale compute; Carbon-X fits on a single GPU, lowering the barrier for any lab.

February 2019

GPT-2 staged release (2019)

OpenAI trained GPT-2, a 1.5-billion-parameter language model, but initially withheld the full model over concerns it could generate realistic fake news and disinformation. The full model was released later that year after safety assessment.

Then

The staged release spawned a public debate about responsible AI model distribution.

Now

Set a precedent for safety-based release decisions that many AI labs now follow.

Why this matters now

Carbon-X's developers excluded viruses from training data to reduce misuse risk — a similar safety decision applied to biological AI, where the same language-model technology could be turned to dangerous ends.

Sources

(18)

+12