--- base_model: Qwen/Qwen2.5-VL-7B-Instruct library_name: peft tags: - mantis - ggufs - vision-language - concise --- # Mantis M1 Flash A concise vision-language model built by ggufs on top of Qwen2.5-VL-7B-Instruct. ## What it does - Reads up to 6 images and answers about them - Direct, short answers — no rambling, no filler - Identity: **Mantis**, built by **ggufs** - Trained on 430 knowledge pairs across physics, chemistry, biology, math, CS, SVG, Three.js, and web dev ## Performance - Base median response: 1,332 chars - Mantis median response: **154 chars (8.6x shorter)** - All answers factually correct ## Thinking mode - `[flash]` prefix: gives direct answers with internal reasoning - `[think]` prefix: shows full reasoning steps before the answer ## Usage ```python from transformers import AutoModelForImageTextToText, AutoProcessor from peft import PeftModel model = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", dtype="bfloat16", device_map="auto") model = PeftModel.from_pretrained(model, "ggufs/mantis-m1-flash-lora") ```