Hugging Face Bio released Carbon-X, a 14.8B-parameter open model for DNA understanding and generation. It accepts raw DNA in chat and answers in English. A hybrid tokenizer handles DNA as 6-mers and text as BPE. It averages 66.7% across five DNA benchmarks without reasoning. That beats every open model tested and matches GPT-6.1 Sol at 1/1,800 the cost. It runs on a single 80 GB GPU. Weights and code are open source. It can serve via vLLM. The model also retains general assistant capabilities. Training used 302B tokens of DNA, biology text, English and instruction data. It was fine-tuned and reinforced with verifiable rewards on DNA tasks.
What changed
Qwen3-14B had no DNA understanding and no DNA tokenizer.
What it unlocks
Analysing DNA sequences directly in a chat interface without extended reasoning.
- 14.8B parameters
- 66.7% average DNA benchmark score
- 302B training tokens
- $300 to analyse 2.4M ClinVar variants
What you need to act on it
- Transformers v5 or later
- trust_remote_code=True for tokenizer
- 80 GB GPU for local inference
Sources