N環球 
Hugging Face Releases Carbon-X, an Open AI Model That Lets You 'Chat With DNA'
Built with researchers in Beijing and Italy, the free open-source model answers plain-English questions about DNA and judges whether mutations are harmful — at a claimed 1/1,800th the cost of comparable commercial models, though its predictions are not yet lab-validated.
Why it matters
This is the first open model that lets ordinary people 'chat with DNA' in plain language: paste in a genetic sequence, ask whether a mutation is harmful, and it explains its reasoning in English — no programming required. Hugging Face says it runs locally on a laptop, so sensitive genetic data never has to leave the user's machine — a key selling point for medical privacy.
Full report
Hugging Face on Thursday (October 22) released Carbon-X, a free open-source model under the Apache 2.0 license. Developed with researchers at Beijing's Zhongguancun Academy, Italy's Telethon Institute of Genetics and Medicine and the University of Naples Federico II, the launch was reported by WIRED.
The model has about 14.8 billion parameters. It starts from Alibaba's open Qwen3-14B and was trained on 302 billion tokens, roughly 55% of them DNA sequences. Its tokenizer reads DNA in six-letter chunks, and it can take in about 49,000 to 50,000 DNA letters at a time — still far short of a whole human genome. It judges whether a mutation looks benign or disease-causing, flags regions that control gene activity, and predicts the next base of a sequence.
Hugging Face claims an average score of 66.7% across five DNA benchmarks, 27 points above the original Qwen3-14B, and says Carbon-X leads every model it tested on the BioReason Coding benchmark and next-base prediction, at about 1/1,800th the per-problem cost of GPT-6.1 Sol. Two of the five benchmarks (FrontierVEP, NextBase) were designed by Hugging Face itself. The model fits on a single 80GB GPU; the team deliberately excluded virus sequences from training to reduce misuse risk.
The developers are upfront about the limits: Carbon-X's predictions have not yet been validated in wet-lab experiments, so they cannot vouch for accuracy or clinical use. 'We've unlocked a new capability in language models, which is fundamentally about being able to chat with DNA,' senior research scientist Lewis Tunstall told WIRED. By contrast, Google DeepMind's AlphaGenome, released last year, outputs predictions in a numeric format and needs more compute.
💬 Comments
No comments yet. Be the first!