Unsloth drops ultra-tiny Qwen 3.5 0.8B quants

This 2-bit quantization push focuses on finding the absolute floor for running multimodal LLMs on legacy hardware rather than general-purpose utility. The UD-IQ2_XXS variant is part of Unsloth Dynamic 2.0, which attempts to maintain coherence even at sub-2-bit levels. At 0.8B parameters, the model is best suited as a high-speed draft model for speculative decoding to accelerate larger 72B+ Qwen variants. Its support for vision in a tiny footprint also makes it a candidate for simple on-device OCR or image classification on microcontrollers. Real-world utility is likely limited to narrow, fine-tuned tasks or agentic "glue" logic where memory footprint is the primary constraint, as the low output quality highlights the diminishing returns of aggressive quantization on tiny models.