Hugging Face releases Carbon-X, an open 14.8B DNA language model that reads genomes in chat
_lewtun · x · 2026-10-26
Hugging Face Bio (with Zhongguancun Academy and TIGEM/University of Naples Federico II) released Carbon-X, an Apache 2.0 14.8B decoder-only model that reads raw DNA in chat and answers in English, predicting mutation pathogenicity, regulatory regions and species.
- Architecture: no separate DNA encoder — starts from Qwen3-14B with a hybrid tokenizer keeping Qwen3's text vocab and turning DNA into 6-base tokens; 302B tokens of DNA+text mid-training, then SFT and RL
- Specs: 40,960-token context, runs on a single 80GB GPU in bfloat16; Q80 GGUF is just 15.8GB
- Performance: 66.7% average across 5 DNA benchmarks (thinking off) vs 67.5% for GPT-6.1 Sol at max effort; leads all tested models on BioReason Coding (74.9%) and NextBase (39.4%), though splice sites score only 29.8% and it's not clinically validated
- Cost: scoring 2.4M ClinVar variants costs $300 vs $32K–$4M on frontier models
Related event: Hugging Face Open-Sources Carbon-X, a 14.8B DNA Language Model(3 posts)→
More from Models
- Logan Kilpatrick: experimental feature latency cut from 24+ hours to 10 minutes — TheMoonMidas · 2026-10-26
- AI conjures a grand 'atemporal cosmic efficiency' theory of everything, and it's gloriously unhinged — aiamblichus · 2026-10-26
- One month since Opus 5.5 release, author shares retrospective — sammcallister · 2026-10-26
- Users lament Opus-class models being blocked from reverse engineering on consumer hardware — nitarshan · 2026-10-26
- Blogger predicts Anthropic will ship a new model pre-IPO to maximize its $2T valuation — kimmonismus · 2026-10-26
- MemTensor's MemT-1.0-Nano trends on Hugging Face, a memory-tuned Qwen3.6-35B-A3B finetune — MemTensor · 2026-10-26