GLM-5.2: Z.ai's Open Model Now Runs on Local Hardware

What GLM-5.2 is

Unsloth has published dynamic GGUF quantizations of GLM-5.2, the new open model from Z.ai, making it possible to run it on local hardware instead of rented cloud capacity. The model has 744 billion parameters, of which 40 billion are active at any moment, and a context window of one million tokens. Unsloth says it reaches state-of-the-art results on long-horizon coding, reasoning and agentic tasks, and that it performs on par with Claude 4.8 Opus, GPT-5.5 and Gemini 3.1 Pro across many benchmarks — a comparison that comes from the quant publisher itself, not from an independent evaluation.

The company credits Z.ai with providing day-zero access to the weights, and the quantized files are published as GLM-5.2-GGUF on Hugging Face.

Quantization: 86% smaller, roughly a fifth less accurate

The interesting part is what happens to accuracy when the model shrinks. In Unsloth's measurements, the dynamic 1-bit version reaches about 76.2% top-1 accuracy while being 86% smaller than the full 1.5 TB model, and the 2-bit version reaches roughly 82% accuracy at 84% smaller. The 4-bit and 5-bit builds (UD-Q4_K_XL and UD-Q5_K_XL) are described as mostly lossless.

Unsloth also warns against reading the top-1 figure as a capability score. A 76% top-1 result does not mean the model answers incorrectly a quarter of the time; it means that on filler and stop words the model's preferred token changes — "I will now create a novel" may become "The novel is below". To back that up, the team published KL-divergence measurements, which track how far the quantized model's output distribution drifts from the baseline.

Memory requirements

The practical question for anyone with a workstation or a Mac is how much memory each quant needs, counting VRAM and system RAM together. Unsloth's table lists 223 GB for 1-bit, 245 GB for 2-bit, 290–360 GB for 3-bit, 372–475 GB for 4-bit, 570 GB for 5-bit and 810 GB for 8-bit. The recommended UD-IQ2_M build takes 239 GB on disk, which fits on a 256 GB unified-memory Mac, and the same quant can run on a single 24 GB GPU together with 256 GB of RAM when MoE offloading is used.

Running it: llama.cpp and Unsloth Studio

The guides cover two paths. With llama.cpp, the model can be downloaded with the hf CLI using the pattern UD-IQ2_M (or UD-IQ1_S for the 1-bit build), then started with llama-cli at temperature 1.0, top-p 0.95 and min-p 0.01; the maximum context window is listed as 1,048,576 tokens. Unsloth Studio, an open-source web UI, offloads to RAM automatically, detects multi-GPU systems and installs with a single shell command.

GLM-5.2 ships with three thinking modes — non-thinking, and two reasoning levels, high and max — selectable through the reasoning_effort parameter or switched off entirely with enable_thinking set to false. Unsloth recommends max thinking for complicated tasks.