Hugging Face releases Carbon-X, a 14.8B open DNA language model

The model was trained in four stages, including embedding warmup, mid-training on 302B tokens, supervised fine-tuning, and reinforcement learning. Hugging Face reports that Carbon-X achieves a 66.7% average across five DNA benchmarks, outperforming other open models. It scores 74.9% on BioReason Coding, leading all tested models, though it trails GPT-6.1 Sol by 0.8 points overall. Performance on splice-site questions remains low at 29.8%, and the model is not clinically validated.

Carbon-X is designed for cost-effective inference, running on a single 80 GB GPU. Hugging Face estimates that scoring 2.4 million ClinVar variants would cost approximately $300 using Carbon-X, compared to $32,000 to $4 million for frontier models. The release includes deployment options via vLLM, Transformers, and llama.cpp, as well as agent tools for integration into existing workflows.