Carbon-X: 14B DNA-language model beats open LMs, 1000x cheaper inference, Apache 2.0
lvwerra · x · 2026-10-22
Hugging Face releases Carbon-X, a 14B model uniting DNA and natural language:
- Beats every open language model tested on DNA benchmarks; within a point of GPT-6.1 Sol at >1,000x lower inference cost.
- Mid-trained on 300B tokens of interleaved DNA and text, followed by SFT (no distillation) and RL.
- New hybrid tokenizer encodes DNA with 3x fewer tokens than BPE.
- Acts as a DNA specialist subagent: Kimi K3 gains 11.1 points when consulting Carbon-X.
- Apache 2.0, runs locally on a laptop via llama.cpp, keeping DNA data on-device.
Related event: Hugging Face Open-Sources Carbon-X, a 14B DNA-Language Model(20 posts)→
More from Models
- Ant Group releases open medical reasoning model Ling-3.0-flash-Sante (124B MoE, MIT) — AdinaYakup · 2026-10-23
- Fastino's GLiDE v1.2 tops HF Decision Index at 71.5, adds multimodal inputs — multimodalart · 2026-10-23
- 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
- Bittensor subnet's REAX 1, fine-tuned from Perplexity Decider, beats its base across all 10 areas — markjeffrey · 2026-10-23
- Kimi K3.1 vs Qwen 4: both build a 3D helicopter with exploded views in ~15 minutes — Scobleizer · 2026-10-23