Hugging Face releases Carbon-X, an open AI model for asking questions about DNA

The model and its limits

According to Hugging Face’s model card, Carbon-X has 14.8 billion parameters and starts from Qwen3-14B. Its developers trained it on DNA and natural-language text, using a tokenizer that reads DNA in six-base groups. They report 302 billion tokens of mid-training, followed by supervised fine-tuning and reinforcement learning.

Hugging Face says the model’s weights fit on a single 80 GB GPU in bfloat16 and documents ways to run it with vLLM and Transformers. That does not extend to a full human genome: WIRED describes inputs of roughly 50,000 base pairs and says users cannot upload full human genomes. The model card says inputs substantially longer than those used in training have not been evaluated.

Predictions are not a diagnosis

Hugging Face reports an average score of 66.7% across five DNA benchmarks with thinking disabled, compared with 67.5% for GPT-6.1 Sol under its comparison settings. The Carbon-X team built two of those benchmarks, FrontierVEP and NextBase. The model card also warns that Carbon-X can miscount bases, confuse coordinates or invent a sequence’s identity or function when context is missing.

More fundamentally, the model card says Carbon-X has not been clinically validated and should not be used to make medical decisions. WIRED reports that its developers have not validated it in wet-lab experiments, so they cannot say how accurate its predictions are.

Many accounts welcomed Hugging Face’s Carbon-X release for advancing open DNA-language modeling and gene research, while others replied with light jokes about its everyday usefulness.

Based on 26 sentiment-bearing replies from 24 accounts across 4 conversations.

Many accounts welcomed Hugging Face’s Carbon-X release for advancing open DNA-language modeling and gene research, while others replied with light jokes about its everyday usefulness.

Based on 26 sentiment-bearing replies from 24 accounts across 4 conversations.