Hugging Face’s Carbon-X Brings DNA Questions to Local Computers

Hugging Face’s Carbon-X lets researchers put DNA sequences into a chat and ask what a mutation might do. Released October 22, the open model can run on local hardware, making unpublished genetic material available to an AI assistant without necessarily sending it to an outside service. Announcement

That convenience depends on which software loads the model. Its main instructions still say a popular local runtime cannot handle its DNA vocabulary. A separate official download page identifies a supported version, while the runtime’s own code history documents the change. Downloading a file successfully can leave a researcher with the wrong sequence representation. Model Files Runtime

Carbon-X has 14.8 billion learned numerical parameters and builds on Alibaba’s Qwen3-14B. Its developers report a 66.7% five-test average, compared with 67.5% for GPT-6.1 Sol. Model

The collaboration includes Hugging Face, Zhongguancun Academy and the Telethon Institute of Genetics and Medicine and the University of Naples Federico II. Research scientist Lewis Tunstall’s October 22 announcement describes an Apache 2.0 release and local use with llama.cpp. WIRED reported that the researchers designed the smaller model partly around restrictions on uploading sensitive patient data to outside AI providers. Report Announcement WIRED

What the five DNA scores measure

Carbon-X leads GPT on coding-region mutations and next-base prediction. GPT leads on the other three tests. Each comparison model appears at its best average reasoning setting; Carbon-X used one sample per question and thinking disabled. Model

Recalculating the unweighted average gives 66.68% for Carbon-X and 67.54% for GPT, a difference of 0.86 percentage points. Every benchmark contributes one fifth, regardless of its sample size or scoring method. This average therefore combines performance on sequence continuation with performance on mutation classification. It does not estimate how often a model will answer an arbitrary genetics question correctly. Model

Performance also varies inside GenomeQA: Carbon-X scores 93.2% on organism classification and 29.8% on splice sites, where cells join pieces of genetic messages. Its developers say exact counting, coordinates and edits can fail. They recommend using ordinary code for precise operations alongside the model. Model Report

The release report describes why training examples required more than fluent explanations. Early attempts at learning from generated answers carried unsupported biological claims into the training material. The team instead worked with domain experts on checks against the supplied sequence, then used task-specific rewards to improve answers. Its training mixed genomes, biology documents, general English and instructions over roughly 300 billion tokens. Report

Earlier DNA-language systems split the job between models. In the original BioReason project, first posted in May 2025, Adibvafa Fallahpour and colleagues paired DNA models with language models and trained them on biological reasoning and mutation tasks. Carbon-X takes another architectural route, placing DNA and text tokens in one decoder. The approaches expose different questions about which information actually drives an answer. BioReason Report

Which local download works

Carbon-X’s DNA vocabulary groups bases in six-letter pieces. That is a different operation from the text tokenizer that turns English into fragments. The main model card recommends Transformers or vLLM and says llama.cpp cannot load its custom Python tokenizer. The official GGUF repository, however, says llama.cpp release b11742 or later supports these files and warns that older builds tokenize DNA incorrectly. Files Model

The runtime change is documented independently of those instructions. A llama.cpp contribution merged October 19 added automatic DNA tagging, treatment of incomplete final groups and explicit sequence markers. In the discussion, a maintainer reported matching the Python reference in 1,256 of 1,258 fuzz cases, with two differences involving unusual Unicode handling. That developer test checks tokenization behavior; it is not a clinical study or an independent reproduction of Carbon-X’s benchmark scores. Runtime

The official compressed downloads range from 9.6 GB to 29.7 GB. Hugging Face recommends the 15.8 GB Q8_0 file, saying it preserves the benchmark scores of the larger version. File size is only part of memory use. Its serving instructions also budget about 0.16 MB of working cache per token and identify a maximum context of 40,960 tokens. Files

Using those published estimates, a full context adds about 6.6 GB of cache: 40,960 multiplied by 0.16 MB. Added to the 15.8 GB download, that is roughly 22.4 GB before other runtime memory. This is a calculation from the instructions, not a measurement on a laptop. A shorter context changes the cache requirement, so the download’s size alone cannot establish which machine will run a particular request. Files

The cost headline uses another specific setup. Hugging Face compares commercial API prices with Carbon-X throughput on an Nvidia H200 priced at $5 an hour. Its estimate of about $300 for 2.4 million uncertain ClinVar variants is an inference-cost projection. It does not include laboratory experiments or the professional review needed to establish what those variants mean. Report

Testing whether answers use the DNA

An October 1 preprint shows why biological chat systems need tests beyond final-answer accuracy. Ada Fang and colleagues examined six models, including BioReason, by changing biological inputs while keeping other information fixed. Shuffling DNA barely changed BioReason’s disease-prediction accuracy. When biological inputs conflicted with accompanying text, BioReason and BioReason-Pro followed the text in 97.9% and 99.7% of cases. Study

Carbon-X was not among those models. The study also found that biological inputs contributed to other systems’ performance, so its result was not a verdict on every DNA-language architecture. Its experiments identify a test that remains relevant to the new release: whether an answer changes appropriately when a sequence changes, rather than simply sounding plausible from the surrounding description. The researchers separately examined 165 questions whose identical wording appeared with different genomes and disease labels. This subset let them check whether models used biological differences when the text alone could not identify the correct label. Study

Clinical interpretation has a separate evidence process. ClinVar, the U.S. government’s archive of genetic variant classifications, records who submitted an assessment, whether criteria were supplied, agreement among submissions and expert-panel review. A model’s prediction does not automatically acquire any of those review statuses. Carbon-X’s own card says the model has not been clinically validated and should not guide medical decisions. ClinVar Model

In the launch thread, Tunstall said the team would also release its full training recipe and data. That would let outside groups examine more of the choices behind the reported results. Announcement