Hugging Face releases Carbon-X, an open 14B model that lets users chat with DNA, close to GPT-6.1 Sol on DNA benchmarks at ~1/1,800 the cost
Confirmed
The takeaway
Hugging Face released Carbon-X, an Apache-2.0 14.8B model built on Qwen3-14B and trained on 300B tokens mixing DNA and English, so users can paste raw DNA into a chat and get explanations of variant effects, regulatory regions and organism groups.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 22 October 2026
Your AI and this story
- GPT-6 Astra175 days after its cutoff
- Claude Opus 5.5114 days after its cutoff
- Gemini 4 Argon205 days after its cutoff
- Grok 4.7144 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 114 days before it.
Key facts
- Model card: 14.8B decoder-only, base Qwen/Qwen3-14B, Apache 2.0; one token stream for DNA and text; DNA tokenised as non-overlapping 6-mers
- Benchmarks (Hugging Face): 66.7% average across five DNA benchmarks with thinking off, vs GPT-6.1 Sol 67.5%, at about 1/1,800 of the inference cost per problem; best of all tested models on coding-variant effects and next-base prediction
- Training (WIRED): continued training on 300B tokens of DNA, English text and paired DNA–description examples from mammals, other vertebrates, invertebrates, plants, fungi and protozoa
- Partners: Zhongguancun Academy (China) and University of Naples Federico II; Lewis Tunstall: ‘We’ve unlocked a new capability in language models, which is fundamentally about being able to chat with DNA’
- Weights: huggingface.co/HuggingFaceBio/Carbon-X (and GGUF)
What happened
Hugging Face’s biology team extended a general LLM’s vocabulary with DNA 6-mers and kept training it on mixed genomic and text data, producing a conversational DNA model small enough to run locally. WIRED notes Hugging Face agreed in September to be acquired by Nvidia.
Why it matters
Genomic models like AlphaGenome and Evo 2 answer in numbers or sequences; Carbon-X answers in plain language, cheaply and openly, which widens access to variant interpretation (with the usual caveats for clinical use).
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 primary, 1 press
Primary
- Hugging Face: Carbon-X model cardhuggingface.co, docs
Press
Changes
- Filed (Techmeme)