THE SHORT VERSION
● Size — 14 billion parameters
● Licence — Apache 2.0 - commercial use permitted
● Context — roughly 50,000 base pairs
● Caveat — predictions have no wet-lab validation
Carbon-X is a 14-billion-parameter open model from Hugging Face that takes plain-English questions about DNA. It ships under Apache 2.0, so commercial use is permitted without negotiation, with a context window of roughly 50,000 base pairs. Hugging Face describes it as approaching GPT-6.1 Sol on internal benchmarks.
Two things make this more interesting than the parameter count suggests. The licence is genuinely permissive, which in biology matters more than in most domains because the organisations that would use it — university labs, small biotechs, agricultural research groups — are exactly the ones without budget for frontier API access. And a 14B model runs on hardware those groups already own.
The caveat is the one Hugging Face states itself: the predictions lack wet-lab validation. Internal benchmarks measure whether the model answers consistently with its training data, not whether the biology is correct. Until someone runs the experiments, Carbon-X is a hypothesis generator rather than an answer engine — which is still useful, provided nobody mistakes the second thing for the first.
Sources
TechStartups