Hugging Face's Carbon-X is an open 14B model you can chat with about DNA

Carbon-X starts from Qwen3-14B. The team added 5,515 DNA tokens to its vocabulary, encoding DNA as non-overlapping 6-mers so each token covers six base pairs, and kept the original BPE vocabulary for text. DNA and English share one token stream in a single decoder, with no separate DNA encoder, and DNA in a prompt is tagged automatically[1].

Training ran in three stages. After a short warm-up that taught only the new DNA embeddings, the full model was trained on 302B tokens: 55% raw DNA from mammals, other vertebrates, invertebrates, plants, fungi, protozoa and prokaryotes, plus English, biology text and instruction data. Supervised fine-tuning and two rounds of reinforcement learning with code-checked rewards followed[1]. The full training data is published as the FineDNA dataset[4], and the recipes are on GitHub[5]. Tunstall told WIRED the team left viruses out of training so the model would not be used for potentially dangerous applications[2].

Across five DNA benchmarks covering variant effects, sequence understanding and next-base prediction, Carbon-X averages 66.7% with thinking disabled. That is 27 points above Qwen3-14B and close to GPT-6.1 Sol's 67.5% at max reasoning, at about 1/1,800 of the inference cost per problem. It beats every model Hugging Face tested on coding-variant effects (74.9%) and next-base prediction (39.4%), though frontier models stay ahead on non-SNV variants and single-base variant effects[1].

The cost gap is the headline for researchers. Hugging Face estimates that analysing the 2.4 million variants of uncertain significance in ClinVar would cost about $300 with Carbon-X, versus $32,000 to $4 million with the frontier models it evaluated[1].

Because researchers often cannot upload identifiable patient data to cloud models, the team deliberately kept Carbon-X small enough to run on their own hardware, and Tunstall said no data is sent to Hugging Face or anyone else[2]. It serves with vLLM on one GPU, and GGUF builds for llama.cpp go down to 9.6 GB, with the 15.8 GB Q8_0 file recommended as matching full-precision scores[3]. The model card also ships an agent tool and a ready-made skill so coding agents like Claude Code and Codex CLI can call Carbon-X when a question involves DNA[1].

Hugging Face is blunt: Carbon-X has not been clinically validated and should not guide medical decisions. It can miscount bases or confuse coordinates, may invent a sequence's identity when context is missing, and remains weak on splice sites (29.8%) and histone marks (34.6%). It reads and explains sequences but was not trained to design new ones[1]. The team told WIRED it has not yet checked the model's predictions against wet-lab experiments, and its training on roughly 50,000-base-pair sequences means users query small parts of a genome at a time rather than whole genomes[2].