GLM 5.2

Z.AI

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 3 September 2026

GLM 5.2 held first place among open-weight models on the Artificial Analysis index until summer 2026; as of 2 September 2026 the index puts its successor GLM 5.3 and Kimi K3 at the top of open-weight models, at 60 points. This mixture-of-experts architecture totals 744 billion parameters and targets reasoning, agentic use and code. Its tendency to reason at length calls for monitoring of output tokens.

Key points: Quality benchmark · 744B mixture of experts · Reasoning and code · Agentic · Open weights.

Architecture

GLM 5.2 is a mixture-of-experts model with 744 billion parameters as announced by Z.AI: 256 routed experts of which 8 are active per token, 78 layers, a native context of 1,048,576 tokens, 1,507 GB of weights published in BF16 (753 billion tensor elements), MIT licence (config.json and Hugging Face repository re-read on 2 September 2026). Only a fraction of the experts activate per token, which contains the compute cost at inference. On 25 August 2026 Z.AI released its successor GLM 5.3, on the same architecture, as FP8 weights (756 GB) under a model-specific GLM-5.3 licence distinct from MIT; GLM 5.3 Flash (320 billion, 18 active, MIT) is a separate model.

Strengths

The model held first place among open-weight models on the Artificial Analysis index until summer 2026; as of 2 September 2026 the index puts GLM 5.3 and Kimi K3 at the top of open-weight models, at 60 points. It targets reasoning, agentic use and code generation.

Use cases

GLM 5.2 serves demanding reasoning and coding tasks. Its tendency to reason at length calls for monitoring of output tokens to control cost.

A public deployment: mmr.qdna.fr has GLM 5.3 and MiniMax M3, drawn in turn, arbitrate its 227 analyses, and publishes the majority vote of the three draws.

Deployment

Served by vLLM on a GPU server, or by quantised llama.cpp. Together with DeepSeek V4, it forms the backbone of the dwarfstar stack.

All models on the platform are interchangeable through a single gateway: switching models amounts to changing one configuration line, with no code rewrite. The priority remains local execution.

Official sources

• zai-org/GLM-5.2, official model card on Hugging Face: licence, parameters, context window and weight format
• zai-org/GLM-5.3, official card of the successor: FP8 weights and GLM-5.3 licence
• Artificial Analysis, model comparison index