Carbon-X: Open 14B DNA Language Model Beats Open Rivals, Runs Locally via llama.cpp
loubnabnl · reddit · 2026-10-22
The HuggingFaceBio team released Carbon-X, a 14B model that unifies DNA and natural language for the first time.
- Beats every open model tested on five DNA benchmarks and matches GPT-6.1 Sol at roughly 1,000x lower cost
- Recipe: midtrained from Qwen3 14B on a DNA + natural language mix using a new hybrid tokenizer, plus RL; training details published in a blog post
- Use cases: query DNA sequences in plain English, predict mutation effects, generate new DNA; a GGUF version enables local inference via llama.cpp
- Motivation: genomics data is highly sensitive and frontier models often refuse routine genomics requests, making open models critical for AI for Biology
More from Models
- Users report free $100/mo Claude API credits just by linking an organization in billing settings — ThePeterMick · 2026-10-22
- Even an RTX PRO 6000 can't replace cloud LLMs, argues Reddit work-usage test — C1rc1es · 2026-10-22
- Users ask whether Codex has been quietly nerfed — Gauri_the_great · 2026-10-22
- Flint int4 hits 122.6 tok/s on Mac mini matching Qwen3.6-35B quality at 3.5x speed — _rockt · 2026-10-22
- "Model at capacity": user complains Astra may be gone as limits push cancellations — Shppo · 2026-10-22
- Bayesian take: LLM benchmark noise means all models are probably tied — remilouf · 2026-10-22