Hugging Face has introduced Carbon-X, a 14B open-source DNA language model built on the Qwen3-14B architecture. The model utilizes a hybrid tokenizer that integrates the standard Qwen3 BPE vocabulary with 5,515 specialized DNA tokens, including 4,096 non-overlapping 6-mers. This design allows the model to process DNA sequences and natural language within a single decoder stream, eliminating the need for a separate DNA encoder.
The model architecture features 40 layers, a hidden size of 5120, and Grouped Query Attention with 8 KV heads. According to Hugging Face, the development team balanced DNA learning and English proficiency by adjusting the learning rate and utilizing a 14B parameter scale. In benchmarks, Carbon-X outperformed the base Qwen3-14B model by 27 points and achieved a 79.3% accuracy rate on IFEval.
Carbon-X is offered at a throughput-based price of $5 per hour on H200 hardware. Hugging Face estimates that processing 2.4 million ClinVar variants costs approximately $300, compared to costs ranging from $32,000 to $4 million for frontier models. The model also functions as a sub-agent, reportedly increasing the DNA index performance of Kimi K3 from 49.4% to 60.5%.
