Hugging Face releases open 14B Carbon-X model for DNA and text
A source-linked abstract prepared once when this article was collected.
Carbon-X is a 14.8B-parameter decoder-only language model trained jointly on DNA sequences and natural-language text, released on Hugging Face under Apache-2.0. It uses a hybrid tokenizer with BPE for text and non-overlapping 6-mers for DNA, and is post-trained and evaluated with thinking disabled. Across five DNA benchmarks the model averages 66.7%, which the source reports as 27 points above its base model Qwen3-14B and ahead of other open models evaluated. The weights reportedly fit on a single 80 GB GPU in bfloat16.
Key takeaways
- Carbon-X averages 66.7% across five DNA benchmarks with thinking disabled, versus 27 points above its base Qwen3-14B.
- DNA and text share one token stream; DNA is encoded as non-overlapping 6-mers, with no separate DNA encoder.
- The source reports the model is close to GPT-6.1 Sol (67.5%) at roughly 1/1,800 of its inference cost per problem.
Why it matters
An open, single-GPU model for DNA reasoning would let researchers analyze sensitive or unpublished sequences on their own infrastructure, rather than sending data to hosted frontier APIs.
Read with care
Source claims not independently checked; benchmark scores and cost estimates are the authors' own, and several benchmarks were built by the authors.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.