Hugging Face's Carbon-X: 14B open model with 300B DNA tokens rivals GPT 6.1 Sol
evijit · x · 2026-10-30
A Hugging Face researcher detailed Carbon-X, a 14B open-source model built by mid-training Qwen3-14B on 300 billion tokens of English and DNA. He argues existing architectures and data can't be pushed much further for science and RL fine-tuning won't help — changing the pretraining mix can. On genomics benchmarks Carbon-X rivals GPT 6.1 Sol at a tiny fraction of the cost, showing academics without closed-lab resources can still compete. He also warns open models are falling behind on many science benchmarks.
More from Models
- DeepSeek's cyber model enters internal testing; observers back keeping it closed — Xianbao_QIAN · 2026-10-30
- Same agent harness, opposite rankings: TB4 benchmark drama exposes eval chaos — lunwang1996 · 2026-10-30
- 90+ post-release benchmarks: Kimi K3 matches GPT-5.5/Opus 4.8 but trails Fable 5 badly — scaling01 · 2026-10-30
- Gemini 4 Carbon is maybe the best model on paper, but Redditors aren't convinced — tjrobertson-seo · 2026-10-30
- User complains Claude Opus has too much "ego", ignores instructions — letandrewcook · 2026-10-30
- Screenshot appears to show GPT-6 as a selectable model in ChatGPT app settings — manosaie · 2026-10-30