The Pulse
Hugging Face Releases Carbon-X, a 14.8B DNA Model
Hugging Face has released Carbon-X, an open 14.8-billion-parameter model that handles DNA sequences and natural-language questions in one conversation. The company reports competitive results on five DNA benchmarks, while warning that the m
AI.info Team · Written with AI · fact-checked ·
Hugging Face released Carbon-X on October 22, 2026, presenting an open model that takes DNA sequences and questions in ordinary language and responds in English. The 14.8-billion-parameter model can assess possible mutation effects, identify biological features in a sequence, classify sequences by organism group and predict subsequent DNA bases. Its model files carry an Apache 2.0 license.
The release gives researchers a downloadable option for combining genomic data and conversational prompts in one system. But Hugging Face describes Carbon-X as a research starting point, not a clinical tool: it has not been clinically validated, can make errors, and should not guide medical decisions.
One token stream for DNA and text
Carbon-X starts from Alibaba’s Qwen3-14B and was further trained on DNA alongside natural-language text. Rather than using a separate DNA encoder, it processes DNA and text through a single decoder and token stream. Its hybrid tokenizer retains Qwen’s text vocabulary and adds DNA tokens that represent six consecutive bases at a time.
The team reports 302 billion tokens of mid-training, followed by supervised fine-tuning and reinforcement learning on DNA tasks. The model’s context window is 40,960 tokens, though its training and post-training examples were shorter. Hugging Face says the weights fit on a single 80GB GPU in bfloat16, allowing researchers to run the model on their own infrastructure rather than send sequences to an external service.
Carbon-X’s benchmark lead comes from the team’s tests
Across five benchmarks covering mutation effects, sequence understanding and next-base prediction, Hugging Face reports an average score of 66.7% with reasoning disabled. The company’s comparison puts GPT-6.1 Sol at 67.5% and says Carbon-X’s inference cost per problem is about one eighteen-hundredth as high. The evaluations are the developers’ own; the model card identifies two benchmarks as drawn from prior research and two as built by the Carbon team.
Results vary by task. Carbon-X scores 74.9% on the coding-variant benchmark, ahead of the listed comparison models, but scores 29.8% on splice-site questions and 34.6% on histone-mark questions in GenomeQA. Those uneven results matter: the overall average does not describe equal performance across different kinds of biological interpretation.
The model card warns against clinical use
“Carbon-X is a starting point for research on connecting DNA and language,” Hugging Face Bio writes in the model card. The team warns that the model may invent a sequence’s identity or function when prompts lack context, and can miscount bases, confuse coordinates or make incorrect edits. It advises pairing the model with other tools for exact sequence operations and checking its biological claims.
Carbon-X can generate DNA continuations, but the team says it was not trained to design new sequences and that quality declines on long generations. The model card also warns that genetic data is sensitive: privacy depends on how users run the model, including whether they use a local download, an inference service or a third-party host. Its release makes the weights available for research and local deployment; it does not establish that the model’s predictions are reliable enough for patient care.