Carbon-X: open 14.8B model takes raw DNA in chat and answers in English
jatingargiitk · reddit · 2026-10-22
Carbon-X is a 14.8B decoder-only model from Hugging Face Bio (Apache 2.0) that accepts raw DNA in a normal chat and replies in English. It shares one token stream for text and DNA using 6-mer tokens and was mid-trained on 302B tokens. HF reports 66.7% average across five DNA benchmarks (thinking disabled) vs 67.5% for GPT-6.1 Sol; weights fit on one 80GB GPU in bfloat16. The model is not clinically validated, and exact sequence operations should still be handled by code. The poster highlights local sequence analysis for sensitive datasets and asks whether gains hold on private variant data.
Related event: Hugging Face Open-Sources Carbon-X, a 14B DNA-Language Model(20 posts)→
More from Models
- Claude Can't Stop Writing Sycophantic Comments, Devs Say Gemini Astra Is Cleaner — ctjlewis · 2026-10-23
- Late Interaction releases a short technical report on its late-interaction retrieval models — lateinteraction · 2026-10-23
- Kimi K3.1 vs Qwen 4: both build a 3D helicopter with exploded views in ~15 minutes — Scobleizer · 2026-10-23
- FreshStack retrieval leaderboard adds Cohere Embed v5, EmbeddingGemma2, Perplexity late-interaction models — beirmug · 2026-10-23
- Together AI and Stanford's Next-Concept Modeling: 10-31x faster than diffusion LMs at equal quality — jiqizhixin · 2026-10-23
- Full-duplex competition will hinge on interaction timing, not just answer quality — kimmonismus · 2026-10-23