Qwen3 4B to 32B in every format we ship: imatrix GGUF (seven tiers each) for llama.cpp, plus FP8, AWQ and GPTQ for vLLM.