Running Qwen3-8B-27B locally with LM Studio on an RTX 4060 (8GB VRAM): 1-bit quantization, thinking-mode repetition issues, and real tokens/s benchmarks.