Open-source DNA model Carbon-X (14.8B) nears frontier scores on genomics at 1/1800th the cost

aigclink · x · 2026-10-25

Hugging Face open-sourced Carbon-X, a 14.8B model unifying DNA and natural language in one Transformer, built on Qwen3-14B with 5,515 DNA tokens (6 bases per token), 302B tokens of continued pretraining, plus SFT and verifiable-reward RL.

Ships with Skills for Claude Code/Codex and tool-calling examples for agent workflows — a low-cost, self-hostable option for genomic database analysis and variant research.

Original post →

More from Models

Models channel →