Hugging Face releases open DNA model Carbon-X, matching GPT-6.1 Sol at 1/1800 the cost
LoubnaBenAllal1 · x · 2026-10-22
Hugging Face's bio team launched Carbon-X, billed as the best open DNA model: they midtrained Qwen3-14B on 300B tokens of DNA, English and interleaved data, using a hybrid tokenizer (BPE for text, 6-mer for DNA).
Key points:
- Across five DNA benchmarks, Carbon-X averages 66.7, beating every open model evaluated (+27 points over Qwen3-14B) and nearly matching GPT-6.1 Sol (67.5) at roughly 1/1800 the cost per problem.
- Strongest on mutation-effect prediction (BioReason Coding 74.9) and DNA sequence completion (NextBase 39.4).
- As a sub-agent, Carbon-X lifts Kimi K3's DNA intelligence index from 49.4% to 60.5% (+11.1pp).
- Midtraining unlocked gains RL alone couldn't: RL directly on Qwen3-14B never beat the 50% chance baseline on FrontierVEP, while starting from the midtrained checkpoint it kept improving.
- The team also shows frontier models increasingly refuse routine genomics as reasoning effort rises: Opus 5.5 drops from 17.7% to 1.1% on next-base prediction, highlighting why open science matters in AI for biology.
More from Research
- IBM's LASAGNA study shows RAG technique gains shrink when combined: +47% predicted becomes +17% realized — ibm · 2026-10-22
- Dev hot take: mechanistic interpretability is 'astrology for AI people' — dejavucoder · 2026-10-22
- eLife 'Landmark' paper: RNNs reveal a mechanistic theory of planning in prefrontal cortex — ShahabBakht · 2026-10-22
- Mathematician Kontorovich: 'All code needs to be Lean formalized like yesterday!' — ludwigABAP · 2026-10-22
- Cohere's CulturalRiddles: 28 LLMs score 79.3% on cross-cultural riddles, but just 22% in some cultures — Cohere_Labs · 2026-10-22
- Berkeley-Stanford preprint: agents that periodically forget beat ones that remember everything — ombre-ai · 2026-10-22