Hugging Face Bio releases Carbon-X, a 14.8B DNA model
Original title: HuggingFaceBio/Carbon-X — new model trending #29 on Hugging Face
The open chat model, started from Qwen3-14B, reports 66.7% on five DNA benchmarks with thinking disabled.
Hugging Face Bio published Carbon-X, a 14.8B-parameter decoder-only model trained jointly on DNA and natural-language text. It answers DNA questions in English, including variant-effect and next-base tasks, using a hybrid tokenizer that encodes DNA as 6-mers. The publisher says it averages 66.7% across five DNA benchmarks, 27 points above Qwen3-14B, and fits on one 80 GB GPU in bfloat16. Weights are released for local serving with Transformers v5 or later and vLLM.
Unconfirmed: Benchmark scores and cost figures are the publisher’s own measurements, not independent results.
- Architecture is 40 layers, hidden size 5120, with a 40,960-token context.
- Mid-training used 302B tokens, followed by supervised fine-tuning and reinforcement learning.
- The lab estimates ClinVar’s 2.4 million uncertain variants would cost about $300 to analyse.
- It reports 66.7% versus 67.5% for GPT-6.1 Sol at roughly 1/1,800 of the inference cost.
- Organizations
- Hugging Face Bio
From the original article
12 min read at huggingface.cotext-generation · 570 downloads · 151 likes · license apache-2. 0
By HuggingFaceBio · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.