It starts from Qwen3-14B and adds 5,515 DNA tokens to the vocabulary, mostly 6-mers, so one token covers six base pairs. DNA and text share one decoder with no separate encoder. Paste a sequence into a message and the tokenizer tags it on its own.
Across five DNA benchmarks the card reports a 66.7% average with thinking off. That's ahead of every open model the team tested and just behind GPT-6.1 Sol at 67.5%, at what Hugging Face estimates is about 1/1,800 of the cost per problem. The card lists its own weak spots: splice sites at 29.8% and histone marks at 34.6%. It has not been clinically validated, and the team says not to use it for medical decisions.
For agents, the repo has an OpenAI-style tool example, an Inspect agent and a carbon-dna skill. The card's advice is to pass sequences by file name instead of pasting them (language models miscopy long sequences) and to do counting and coordinates in code.
One correction to some early coverage: the card says llama.cpp and other GGUF runtimes are not supported, because the DNA tokenizer is custom Python code. Use vLLM or Transformers v5.