Z.ai’s GLM-5.2 packs 744 billion parameters with 40 billion active ones and a 1 million-token context window. You can run it locally via Unsloth Dynamic GGUFs in quantized forms: the 2-bit dynamic UD-IQ2_M uses 239 GB on disk, fits a 24 GB GPU plus 256 GB of RAM, and hits about 82 percent top-1 accuracy while being 84 percent smaller than the full 1.5 TB model. The 1-bit version drops to ~76.2 percent accuracy and needs 223 GB total memory. For top performance on out-of-distribution tasks, dynamic 4-bit or 5-bit quantizations keep almost all original capacity with minimal KLD loss. GLM-5.2 has three inference modes—non-thinking, high-thinking, and max-thinking—and defaults to reasoning on. You toggle thinking or reasoning effort via command-line flags in Unsloth or llama.cpp. Standard sampling settings work well: temperature 1.0 and top_p around 0.95–1.0. Always leave some headroom so your system RAM plus VRAM exceeds the quant file size by at least 10–20 percent. Unsloth Studio offers a web UI that auto-offloads across RAM and GPUs. Install via pip or curl, launch on localhost (or secure Cloudflare tunnel), then search “GLM-5.2” and pick your quant. The UI sets inference params automatically and provides built-in tools for code execution, web search, and parameter tuning. For a lighter, scriptable setup, clone llama.cpp with GGML_CUDA on or off, point it at the GGUF files, and run conversation mode directly. You can also download models via Hugging Face Hub or manually for speed. Quantization quality comes from dynamic bit allocation: important layers stay in higher precision, unimportant ones drop bits. Mean KLD tests show 4-bit and 5-bit quants are virtually lossless; 1-bit and 2-bit remain useful for most tasks, trading only 18–24 percent accuracy for a four- to six-fold reduction in size. Top-1 accuracy reflects greedy decoding alignment with the unquantized model—your answers stay correct, just a bit more varied in word choice.