Carbon-X
Hugging Face released Carbon-X, a 14.8B open DNA language model that reads raw DNA in a chat and answers in English, under Apache 2.0.
Carbon-X starts from Qwen3-14B and puts DNA and text in one token stream with a hybrid tokenizer that adds 5,515 DNA tokens, with no separate DNA encoder. It was trained in four stages: embedding warmup, mid-training on 302B tokens, SFT and RL with GRPO. It averages 66.7% on five DNA benchmarks, against 67.5% for GPT-6.1 Sol, according to Hugging Face's own evaluation.
- Date
- Sunday 25 October 2026
- Lab
- Hugging Face
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | 14.8B dense, 40,960-token context | company |
| Average on 5 DNA benchmarks | 66.7% GPT-6.1 Sol (Max) 67.5%, Opus 5.5 (Max) 60.9% | company |
| BioReason Coding | 74.9% GPT-6.1 Sol 71.0% | company |
| GenomeQA splice-site questions | 29.8% weakest result | company |
Built with Zhongguancun Academy and TIGEM / University of Naples Federico II; not clinically validated. Release date taken from the coverage date.
Sources
This record was partly confirmed: some claims could not be checked on 25 October 2026. How we check