Hugging Face Open-Sources Carbon-X, a DNA Model You Can Chat With

What happened

Carbon-X was developed by Hugging Face with researchers from Zhongguancun Academy and the University of Naples Federico II, starting from Qwen3-14B. A hybrid tokenizer cuts DNA into 6-base chunks (6-mers), adding 5,515 DNA tokens to the Qwen3 vocabulary so DNA and text share one token stream and one decoder, with no separate DNA encoder; training ran 300B tokens of mid-training followed by supervised fine-tuning and reinforcement learning with verifiable rewards.

Per the model card, Carbon-X can judge whether mutations are benign or pathogenic, identify regions controlling gene activity, classify sequences by organism group, predict the next bases and explain its reasoning in natural language; it reads about 50,000 base pairs per query with a 40,960-token context window. The team positions it apart from DeepMind's AlphaGenome (numeric predictions) and the Arc Institute's Evo 2 (whole-genome generation): Carbon-X does conversational genomic analysis.

Local execution is the headline: weights fit on a single 80GB GPU in bfloat16, and the launch demo ran on a MacBook; senior research scientist Lewis Tunstall said the team deliberately kept the model small so labs need not upload identifiable patient genomic data to frontier models. Thomas Wolf said on X that it scores 67% on their DNA benchmarks vs 61% for Claude Opus 5.5 and 50% for Gemini 4 Argon, within a point of GPT-6.1 Sol at roughly 1,000x lower cost (company-reported). The developers caution that predictions have not been validated in wet-lab experiments, so accuracy is unconfirmed.

Key facts

1. Released Oct 22, 2026; 14.8B-parameter decoder-only model, Apache 2.0, weights on the Hugging Face Hub
2. Architecture: from Qwen3-14B; hybrid tokenizer (Qwen3 BPE + 5,515 DNA tokens, one token per 6 base pairs); 40,960-token context; ~50,000 base pairs per query
3. Training: ~300B tokens (DNA + English + paired biological descriptions) across humans, vertebrates, invertebrates, plants, fungi and protozoa; viruses deliberately excluded
4. Capabilities (per model card): benign/pathogenic mutation calls, regulatory-region identification, organism-group classification, next-base prediction, natural-language reasoning explanations
5. Local runs: fits one 80GB GPU in bfloat16, demoed on a MacBook; genomic data never touches Hugging Face or any company server
6. Benchmarks (company-reported): 67% on DNA benchmarks vs 61% for Claude Opus 5.5 and 50% for Gemini 4 Argon, within a point of GPT-6.1 Sol; predictions not wet-lab validated