Hugging Face open-sources Carbon DNA foundation models with full training recipes
LoubnaBenAllal1 · x · 2026-11-01
Hugging Face released Carbon, a family of fully open genomic foundation models — data, weights, and complete training recipes all public under Apache 2.0.
- The main lineup spans Carbon-500M / 3B / 8B, with zero-shot DNA evals, finetuning recipes, and embedding analysis; the tech report is on bioRxiv.
- Carbon-X adds RL with verifiable rewards on top of SFT and outperforms every open language model on DNA understanding evals, backed by a curated 200k-sample SFT set covering variant effect prediction to sequence QA.
- A joint effort between HF, Zhongguancun Academy, and TIGEM/University of Naples Federico II; code is on GitHub.
Related event: Hugging Face open-sources Carbon DNA foundation models(2 posts)→
More from Research
- If your AI scores 100%, your eval is wrong — an AI safety researcher's eval playbook — AI Engineer · 2026-11-01
- Control Vectors experiment: letting an agent manage its own activations — sloppenheimer · 2026-11-01
- FlashRSI open-sources RSI research: 12 algorithms, 30+ benchmarks, OOD gains still elusive — HuaxiuYaoML · 2026-11-01
- Claude designed scorpion toxin binders: 17 of 211 lab-tested designs bound — shoumikgoswami · 2026-11-01
- Carbon-X: midtrained Qwen3-14B beats every open LLM on DNA benchmarks — LoubnaBenAllal1 · 2026-11-01
- RL fails without midtraining: Qwen3-14B stays at chance on FrontierVEP — LoubnaBenAllal1 · 2026-11-01