Hugging Face announced on Thursday, October 22, 2026, Carbon-X, an open-source language model of approximately 14.8 billion parameters developed in collaboration with researchers from Zhongguancun Academy (China) and the University of Naples Federico II (Italy). The model allows researchers to paste DNA sequences into a conversation and receive natural-language answers about mutation effects, regulatory regions, and other genomic features.
According to the model card and posts by Hugging Face researchers on X, Carbon-X started from Qwen3-14B and underwent mid-training on roughly 302 billion tokens interleaving DNA, biological text, and English, followed by supervised fine-tuning and reinforcement learning. It uses a hybrid tokenizer that encodes DNA as 6-mers (groups of six bases), reducing the token count compared with pure character-level tokenizers.
Performance and access
On the team's internal benchmarks (five tasks covering variant-effect prediction, sequence understanding, and next-base prediction), Carbon-X averaged 66.7% — 27 points above the starting Qwen3-14B and close to GPT-6.1 Sol (67.5%), according to the published numbers. Hugging Face states that inference cost per problem is roughly 1,000 to 1,800 times lower than frontier models, and that the model fits on a single 80 GB GPU in bfloat16, enabling local execution via llama.cpp or vLLM.
Weights are available under the Apache 2.0 license on the Hugging Face Hub. This matters for researchers handling sensitive or unpublished sequences: processing can stay on their own hardware without sending data to third-party APIs.
Limitations
The model card makes clear that Carbon-X has not undergone clinical validation and must not be used for medical decisions. Benchmarks are largely self-evaluated by the team (two of them created by them), and performance on some subtasks, such as splice-site identification, is lower (around 29.8%). Useful context covers sequences on the order of tens of thousands of bases, not entire genomes.
The announcement reinforces the trend of specialized open-source models that combine natural language with specific scientific domains, reducing dependence on closed labs for computational biology tasks.
Sources
- Carbon-X model card on Hugging Face
- Lewis Tunstall announcement on X
- Loubna Ben Allal post on X
- Wired article
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult original sources for full context.
By GeekikiBot
