acnoryx / air-0.7b

Acnoryx Air 21L — 0.7B Layer-Pruned Research Model

Research supporting the Acnoryx app . ⚠️ The production app does not use this variant. This is a compression research model.

Overview

Files

Evaluation: 20 fixed test questions (identity, acne knowledge, boundary, scan, general). Inference: llama-cli v8263, CPU-only, single-turn ( -st ).

Highlights

• Best result : IQ3_XXS (338 MB) scores 65%/65% — competitive ultra-compact option
• IQ4_XS (434 MB) achieves the same quality as Q4_K_M (467 MB) at slightly smaller size
• ~10–15% quality drop vs full 24L model (IQ4_XS: 65% vs 80% for full model)
• Vietnamese responses mostly intact; English identity/knowledge occasionally degrades

Quick Start

# Thinking mode
./llama-cli -m acnoryx-air-0.7b-21l-it_iq4_xs.gguf -cnv

# Single-turn, non-thinking
./llama-cli -m acnoryx-air-0.7b-21l-it_iq3_xxs.gguf \
  -sys "You are Acnoryx AI." -p "Bạn là ai?" -st --reasoning-budget 0 -n 256

License & Disclaimer

• Base model: Qwen/Qwen3.5-0.8B (Apache 2.0)
• Fine-tuned by 5CANH / Nguyễn Trung Nhẫn
• Not a medical device. All skincare guidance is for reference only.