Carbon-X Turns DNA Into a Conversation With Open AI

Hugging Face has released Carbon-X, a 14.8 billion-parameter model that lets users ask questions about DNA in plain English. The open system reads raw genetic sequences, predicts biological outcomes, and answers in a chat, bringing DNA analysis and language generation into one tool.

Lewis Tunstall describes Carbon-X as a “first of its kind” model that lets researchers “chat with DNA.” That promise reaches from mutation analysis to drug and gene-therapy research, but the model also arrives with a clear warning: its predictions have not been validated through real laboratory experiments.

One Model, One Stream for DNA and Text

Carbon-X is a decoder-only model from Hugging Face Bio, built with Zhongguancun Academy and TIGEM / University of Naples Federico II. The team started with Qwen3-14B and expanded its vocabulary so the same decoder can process both natural language and DNA.

Earlier work such as BioReason used a separate DNA encoder. Carbon-X takes a different path: DNA and text share one token stream inside one decoder, with no separate DNA encoder. Its hybrid tokenizer keeps Qwen3’s BPE vocabulary for text and adds 5,515 DNA tokens.

• 4,096 non-overlapping 6-mers, with each token covering 6 base pairs
• 1,364 tail tokens
• 3 markers for the DNA token system
• Automatic tagging for DNA runs containing 6 or more A/C/G/T bases

The backbone uses a standard Qwen3 stack with 40 layers, a hidden size of 5120, and grouped-query attention with 8 KV heads. Carbon-X has a 40,960-token context and comes as a dense model with no mixture-of-experts system.

That design gives users a single conversational interface for several biology tasks. Carbon-X can predict whether mutations are benign or pathogenic, flag regulatory regions, classify organisms, predict the next bases, and answer questions about how DNA changes may affect genetic activity and health.

Training DNA Skills Without Losing Language

Carbon-X training ran through four stages, according to the model card. The first stage used about 10 billion tokens with the transformer frozen, allowing only the new DNA embeddings to learn. Mid-training then processed 302 billion tokens with an 8,192-token sequence length.

The mid-training mixture contained 55% raw DNA, 20% English, 15% biology text, and 10% instruction data. The DNA sequences came from mammals, including humans, along with other vertebrates, invertebrates, plants, fungi, and protozoa. The team did not train the model with viruses because it did not want Carbon-X used for potentially dangerous applications.

Supervised fine-tuning covered 3 epochs across 20 dataset subsets. Domain experts helped build deterministic verifiers that checked the responses. Reinforcement learning then used asynchronous GRPO with verifiable rewards over 2 rounds through TRL and vLLM.

The hardest training control was the learning rate. High rates helped Carbon-X learn DNA but caused it to forget English, while a lower rate, a 14B model, and a long training run fixed that trade-off. The model was trained to process sequences of approximately 50,000 base pairs, so users can consult only small parts of a genome at a time.

Carbon-X also keeps general language skills, recording 79.3% prompt-level strict accuracy on IFEval. It runs with thinking disabled, and Hugging Face evaluated it under that setting. Users should keep enable_thinking=False because the model was post-trained and tested that way.

Strong Results, Low Cost, and Real Limits

Hugging Face reports a 66.7% average score across 5 DNA benchmarks, with 74.9% on BioReason Coding and 29.8% on GenomeQA splice-site questions. Carbon-X sits 27 points above Qwen3-14B and 0.8 points below GPT-6.1 Sol, according to the company’s self-reported results.

It leads on coding-variant effects and next-base prediction, while frontier models still win on complex variants and single-base variants. The comparison also depends on refusals: Opus 5.5 falls from 17.7% at low effort to 1.1% at maximum effort on NextBase, as refusals rise, while Gemini 4 Argon drops from 58% to 38% on FrontierVEP as effort increases.

The model’s open Apache 2.0 weights let laboratories keep sensitive sequences on their own GPUs. Carbon-X can run on one 80 GB GPU, and Hugging Face says it can run on a laptop. No data is sent to Hugging Face or another company when researchers use the model on their own hardware.

Cost is another striking part of the release. Hugging Face prices Carbon-X at measured throughput on an H200 at $5 per hour and reports about 1/1,800 of GPT-6.1 Sol’s inference cost per problem. Scoring 2.4 million ClinVar variants of uncertain significance would cost about $300 with Carbon-X, compared with $32,000 to $4 million on frontier models.

Carbon-X can also serve as a sub-agent. When Kimi K3 consulted the model, its DNA index rose from 49.4% to 60.5%. The release includes an OpenAI-style tool, an Inspect agent, and a carbon-dna skill for coding agents, opening a path toward larger systems that call Carbon-X for specialized biology work.

Developers can use vLLM, Transformers, or llama.cpp to run the model, with Transformers requiring version 5 or later and trust_remote_code=True for the DNA tokenizer. The official Carbon-X-GGUF repository requires llama.cpp b11742 or newer because older builds tokenize DNA incorrectly.

Carbon-X does not turn predictions into proven biological facts, and its developers have not claimed laboratory accuracy. Its importance comes from making DNA analysis more accessible, private, and affordable while keeping the limits visible. As researchers test the model against real experiments, the conversation between language models and biology is moving from a striking idea toward a tool scientists can examine for themselves.

Stay connected via Google News

Follow us for the latest AI News