GLM-5.2

Updated
03.07.2026
Thinking
Reasoning
Code
Run

Run GLM-5.2, a large MoE model, locally with Atomic Chat. Private, offline reasoning and coding with no cloud and no limits. Free.

huggingface-cli download zai-org/GLM-5.2
from transformers import AutoModel
model = AutoModel.from_pretrained("zai-org/GLM-5.2")
More models
View all

  
  

NameSize / UsageContextInput
GLM-5.1
198KText
GLM-4.7-Flash
128KText
At a glance
  • License: Mit
  • Context length: 1M tokens
  • Languages: en, zh
  • Minimum hardware: ~421 GB VRAM
  • Strengths: reasoning, coding and on-device inference
Overview
GLM-5.2 is a 753.3B-parameter open-weight language model from zai-org (Z.ai / Zhipu AI).
  • License: Mit
  • Context length: 1M tokens
  • Languages: en, zh
  • Minimum hardware: ~421 GB VRAM
  • Strengths: reasoning, coding and on-device inference
Highlights
  • Solid 1M context — a 1M-token window that stably sustains long-horizon work.
  • Advanced coding with flexible effort — stronger coding with multiple thinking-effort levels to balance quality and latency.
  • Improved architecture — IndexShare reuses one indexer per four sparse-attention layers, cutting per-token FLOPs 2.9x at 1M context.
  • Pure open — an MIT license with no regional limits.
What it is good at
GLM-5.2 leans toward agentic coding and long-horizon work. These are the tasks it fits.
  • Code generation and review — it writes implementations from a spec and reasons across multiple files, and ranks near the top of open-weight coding benchmarks like SWE-bench Pro and Terminal-Bench.
  • Long-document reasoning — the 1,048,576-token window holds an entire codebase or a long technical document in one pass, so you can ask questions across the whole thing.
  • Step-by-step problem solving — the thinking capability lets it work through multi-step logic and agentic tasks before committing to an answer.
Running it locally
At 753.3B parameters this is a heavy model. Full BF16 weights need hundreds of gigabytes, so most local users run a quantized build: a 2-bit dynamic quant lands around 241 GB and fits a 256 GB unified-memory Mac or a 24 GB GPU paired with 256 GB of system RAM and MoE offloading. The full 1,048,576-token context is available if your memory can hold the KV cache.
huggingface-cli download zai-org/GLM-5.2
From there you can serve it with vLLM or run a GGUF quant through llama.cpp, or load it in Atomic Chat with one click and start a private offline session.
License
GLM-5.2 ships under the MIT license. That permits commercial use, modification, redistribution, and fine-tuning the weights on your own code or domain data, as long as the license notice stays attached.
Desktop



macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download
Frequently asked questions
GLM-5.2 is a 753.3B-parameter open-weight model from zai-org (Z.ai), built on a mixture-of-experts architecture and released under the MIT license. It targets coding, reasoning, and agentic tasks, supports English and Chinese, and has a 1,048,576-token context window. The weights are public on Hugging Face, so you can run it yourself instead of through a hosted API.
The full BF16 weights need hundreds of gigabytes of memory, so most people run a quantized build. A 2-bit dynamic quant is around 241 GB and fits a 256 GB unified-memory Mac, or a single 24 GB GPU combined with 256 GB of system RAM using MoE offloading. On consumer hardware with low-bit quants, expect roughly 3 to 9 tokens per second.
Yes. GLM-5.2 is released under the MIT license, which makes the weights free to download, run, modify, and use commercially. Running it locally through Atomic Chat has no per-token cost since the model executes on your own machine. You only pay for the hardware and electricity.
Yes. Once the weights are downloaded, GLM-5.2 runs entirely on your device with no internet connection required. Nothing is sent to an external server, so your prompts and code stay local. Atomic Chat loads the model on-device for this kind of private, offline use.
Its strongest area is coding and agentic engineering, where it ranks among the top open-weight models on benchmarks like SWE-bench Pro and Terminal-Bench. The 1,048,576-token context also makes it good for reasoning over a full codebase or a long technical document. Its thinking capability suits multi-step problems that need the model to work through logic before answering.