GLM 5.3

Z.AI

Open-weight model · runs on-premises · Published on 5 July 2026 · Updated on 8 October 2026

Updated on 8 October 2026: Artificial Analysis index v4.3.2 re-read on 8 October 2026 (GLM 5.3 at 44.8, ahead of Kimi K3 at 43.6).

Updated on 26 September 2026: Artificial Analysis index re-read in version 4.3.2 on 26 September 2026 (recalibrated scale: GLM 5.3 at 45, second open model behind MiMo-V2.6-Pro at 46).

GLM 5.3 ranks second among open-weight models on the Artificial Analysis index v4.3.2 (45 points, behind MiMo-V2.6-Pro at 46; Kimi K3 at 44; read on 26 September 2026). This mixture-of-experts architecture totals 744 billion parameters for the Deep version and 320 billion (18 billion active) for the Flash version. It targets reasoning, agentic work and code. Its tendency to reason at length calls for monitoring of output tokens.

Key points: Quality benchmark · 45 points Artificial Analysis v4.3.2 · 744B and 320B mixture of experts · Reasoning and code · Agentic · Open weights.

Architecture

GLM 5.3 is a mixture-of-experts model totaling 744 billion parameters as announced by Z.AI for the Deep version: 256 routed experts of which 8 are active per token, 78 layers, extended native context, weights published in FP8 (756 GB) under the GLM-5.3 licence. Only a fraction of the experts activate per token, which contains compute cost at inference. The GLM 5.3 Flash version totals 320 billion parameters (18 billion active per token) under the MIT licence.

Strengths

The model ranks second among open-weight models on the Artificial Analysis intelligence index, at 45 points on version 4.3.2 of 26 September 2026 (index v4.3.2 of 8 October 2026: 44.8), behind MiMo-V2.6-Pro (46) and ahead of Kimi K3 (44). It targets high-level reasoning, agentic orchestration and code generation.

Use cases

GLM 5.3 serves demanding reasoning, analytical 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 572 analyses, and publishes the majority vote of the three draws.

Deployment

Served by vLLM on a GPU server, or by quantised llama.cpp.

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-5.3, official model card on Hugging Face: FP8 weights and GLM-5.3 licence
• zai-org/GLM-5.3-Flash, Flash version: 320 billion parameters, 18 billion active, MIT licence
• Artificial Analysis, model comparison index

Which machine can run GLM 5.3?

GLM 5.3 totals 744 billion parameters for the Deep version (around 428 GB in NVFP4) and 320 billion for the Flash version (around 184 GB in NVFP4, fitting in 512 GB unified memory on Mac Studio Ultra). 7 catalogue platforms have enough memory.

See also GLM 5.3 in NVFP4 and the full sizing matrix.