How to Run GLM-5.2 Locally for AI Coding (2026 Guide)

What Is GLM-5.2 and Why Developers Care

GLM-5.2 is an open-weight large language model from Z.ai (Zhipu AI), a Chinese lab. It shipped to Coding Plan subscribers on June 13, 2026, and the open weights dropped three days later on Hugging Face and ModelScope. It is a sparse Mixture-of-Experts model with 744B total parameters and about 40B active per token, a 1,000,000-token context window, and up to 128K output tokens. The license is MIT, so you can use and deploy it commercially without restriction.

The reason it matters is simple: it became the most capable open-weight model on the Artificial Analysis Intelligence Index the week it launched, the first time an openly licensed model has felt close to the frontier for agentic coding. Nathan Lambert called it the first open-weight model that feels right in coding harnesses as a general agent. That claim is what triggered the wave of "can I finally replace Claude with a local model" threads on Hacker News in June 2026.

Cost is the other half of the story. The hosted GLM-5.2 API runs at roughly $1.40 per 1M input and $4.40 per 1M output tokens, about one-sixth the price of GPT-5.5 ($5/$30) or Claude Opus 4.8 ($5/$25), according to Artificial Analysis. When the API is already that cheap, the case for running it locally comes down to privacy, offline use, and very high volume rather than raw savings - a point I keep coming back to in the verdict section.

What Hardware Do You Need to Run GLM-5.2 Locally?

The first thing to internalize is that MoE saves compute, not memory. Even though only ~40B parameters activate per token, all 744B weights must sit in RAM or VRAM at once, plus headroom for the KV cache (which is large at long context). Offloading inactive experts to system RAM saves on compute throughput, not on the total memory footprint. There is no trick that lets you run this model in 64GB.

Throughput is sobering. Real-world numbers people posted on the Hacker News "how to run locally" thread range from about 1 token per second on a CPU-only 12-core box to roughly 6 tokens per second on dual RTX 3090s with offload. Cloud endpoints serve 80 to 200 tokens per second. A useful rule of thumb from that thread: decode speed tracks active_weights_GB / memory_bandwidth, which is why a Mac's ~615 GB/s unified memory beats a DGX Spark's ~274 GB/s despite the marketing.

How to Run GLM-5.2 Locally (Step by Step)

The primary local path is llama.cpp with an Unsloth GGUF. On a Mac, LM Studio wraps the same engine in a GUI and is the easiest on-ramp. Below is the llama.cpp route, which gives you the most control and the OpenAI-compatible server you need for coding agents.

Step 1: Build llama.cpp

# Build with CUDA (use -DGGML_METAL=ON on Apple Silicon instead)
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build \
  -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j \
  --clean-first --target llama-cli llama-server

Step 2: Download a quant that fits your memory

# 2-bit (~239GB) - fits a 256GB Mac Studio
hf download unsloth/GLM-5.2-GGUF \
  --local-dir unsloth/GLM-5.2-GGUF \
  --include "*UD-IQ2_M*"

# Prefer Q4_K_XL (~467GB) if you have the memory for coding work
# --include "*UD-Q4_K_XL*"

Step 3: Run it with Unsloth's recommended sampling

Z.ai and Unsloth recommend temperature 1.0, top_p 0.95, and min_p 0.01. For long-context coding sessions, quantize the KV cache to stretch how far the 1M window goes before you run out of memory.

./llama.cpp/build/bin/llama-cli \
  --model unsloth/GLM-5.2-GGUF/UD-IQ2_M/GLM-5.2-UD-IQ2_M-00001-of-00006.gguf \
  --temp 1.0 --top-p 0.95 --min-p 0.01 \
  --cache-type-k q4_1 --cache-type-v q4_1 \
  --chat-template-kwargs '{"reasoning_effort":"max"}'

Should You Actually Run GLM-5.2 Locally?

Run GLM-5.2 locally when you have a real reason that the cloud cannot satisfy: a privacy or compliance requirement that data never leaves your machines, genuinely offline work, very high-volume batch agentic jobs where per-token cost dominates, or you already own a 256GB-plus Mac or a multi-GPU server that would otherwise sit idle. In those cases an MIT-licensed model this capable is a genuine option that did not exist a year ago.

Stay on the cloud when you want interactive speed (single-digit tokens per second is painful for tight edit loops), when the task involves hard multi-file reasoning where the frontier models still have an edge, or when you simply do not have the memory. And here is the part that surprises people: the hosted GLM-5.2 API at $1.40/$4.40 is often the saner middle ground - an open model running on someone else's hardware, far cheaper than GPT-5.5 or Opus, with none of the memory headache. Remember the verbosity tax too: at around 42K tokens per task, GLM-5.2 spends tokens freely, which erodes both local speed and any API savings.

If you are evaluating open-weight options more broadly, it is worth comparing GLM-5.2 against the other models I have covered: Qwen Code for a terminal-native open coding agent, Gemma 4 for genuinely laptop-sized local models, and Apple Core AI for on-device inference on Apple Silicon. GLM-5.2 is the heavyweight of that group: the most capable, and by far the most demanding to host.