Needle targets on-device scenarios, with model sizes ranging from 8 to 29 MB; it accomplishes tool calling, structured extraction, and text embedding at 2 bits per weight, and the inference engine is under 1 MB.