Hugging Face Biology Research releases open 14.8B Carbon-X model for DNA and text

Hugging Face Biology Research has released Carbon-X, an open model with 14.8 billion parameters. It was trained on DNA sequences and ordinary text together and is licensed under Apache-2.0. Users can paste raw genetic code into a chat and ask in English whether a mutation is benign or pathogenic, which regions control gene activity, which organism group a sequence belongs to, or what the next bases are. The model card says replies can mix DNA and prose.

DNA and text share one token stream inside a single decoder, with no separate encoder for sequences, according to the model card. A hybrid tokenizer keeps a byte-pair-encoding vocabulary for text and splits DNA into non-overlapping six-letter chunks, so each DNA token covers six base pairs. Runs of at least six A, C, G or T bases are wrapped automatically, so users don't have to tag them. Any six-letter chunk containing a character such as N becomes a single unknown token, which is why the card advises filtering to canonical bases first if full resolution matters.

The model starts from Qwen3-14B. Its creators first added the DNA tokens with randomly initialised embeddings and trained only those for about 10B tokens. They then trained the whole model on 302B tokens of mid-training. In that mixture, raw DNA makes up 55%, English 20% and biology text 15%. The card reports a tradeoff it found in ablations: a low learning rate preserves English but leaves DNA performance at chance, while a high one learns DNA but forgets English. Supervised fine-tuning and reinforcement learning with verifiable rewards on DNA tasks follow.

The headline number is a 66.7% average across five DNA benchmarks with thinking disabled. The card says this is 27 points above Qwen3-14B and ahead of every other open model it evaluated. GPT-6.1 Sol scores 67.5%, but at its maximum reasoning setting. The per-benchmark table shows an uneven split. Carbon-X leads on BioReason Coding (74.9 against GPT-6.1 Sol's 71.0) and NextBase (39.4 against 32.9). GPT-6.1 Sol leads on BioReason Non-SNV (91.7 against 85.1), GenomeQA (67.2 against 63.6) and FrontierVEP (74.9 against 70.4).

Two of the five benchmarks were built by the creators themselves: FrontierVEP and NextBase. BioReason and GenomeQA come from the literature. NextBase measures the fraction of bases predicted correctly across 600 sequences, and Carbon-X's 39.4% is the best score in that table. The GenomeQA breakdown is also uneven. Enhancer versus promoter classification reaches 70.0%, while splice sites score 29.8% and histone marks 34.6%. The card says training data was checked against the benchmarks for overlap in three ways, including whole-span DNA overlap and sequence homology search.

The cost claim rests on a per-problem comparison. The card says Carbon-X's inference cost is about 1/1,800 that of GPT-6.1 Sol per problem. It estimates that analysing the 2.4 million variants of uncertain significance in ClinVar would cost about $300, against $32,000 to $4 million for the frontier models it evaluated. Other models are priced at standard API rates, while Carbon-X's figure comes from measured throughput per H200-hour at $5 an hour. Because Carbon-X answers directly without extended reasoning, each problem also generates few output tokens.

The weights fit on a single 80 GB GPU in bfloat16. The card presents this as the reason researchers can keep sensitive or unpublished sequences on their own infrastructure. It also describes using Carbon-X as a tool for other agents. Agents often answer DNA questions themselves rather than calling the tool, so users who want Carbon-X's view must say so explicitly. The card recommends passing sequences by reference rather than having a language model copy them. A ready-made skill for Claude Code and Codex CLI is included, and GGUF quantizations exist for llama.cpp.

The Carbon project is a collaboration between Hugging Face, Zhongguancun Academy and TIGEM/University of Naples Federico II, according to the GitHub repository. The repository links to a Carbon technical report on bioRxiv and a separate Carbon-X blog post. The model card's limitations section breaks off after the words 'It has not', so the document as available doesn't finish its limitations statement.