GLM-4.5
Open-source mixture of experts model for agents and reasoning
Zhipu AI released GLM-4.5 in August 2025 as an open-source foundation model built for intelligent agents. It combines reasoning, coding, and tool use in a single system with 355 billion parameters.
- 355 billion total parameters with 32 billion active
- Trained on 23 trillion tokens of data
- Scores 91.0% on AIME 24 reasoning benchmark
- Released under MIT open-source licence
- Supports 128K token context window
What it is
GLM-4.5 is a large language model created by Zhipu AI to handle complex agent tasks. It launched in August 2025 with 355 billion total parameters, though only 32 billion are active during any single step. This design allows it to balance high capability with efficient computation. The model focuses on three core areas: acting as an autonomous agent, solving difficult reasoning problems, and writing code. A smaller variant called GLM-4.5-Air offers similar features with 106 billion total parameters for users who need less computational overhead.
How it came about
The GLM-4.5 Team at Zhipu AI developed this model to advance research in agentic artificial intelligence. They published their technical report in August 2025, detailing the training process and evaluation results. The team released both the full model and the Air version to the public simultaneously. This release included base models, hybrid reasoning versions, and FP8 quantised variants. The goal was to provide researchers and developers with powerful tools that could be used commercially without restrictive licensing barriers.
Inside the model
This model uses a Mixture of Experts architecture, which means it contains many specialised sub-networks but only activates a small fraction for each input. It was trained on 23 trillion tokens of data through multiple stages. The training process included expert model iteration and reinforcement learning to improve its performance. The system supports a context window of 128,000 tokens, allowing it to process very long documents. It employs a hybrid reasoning method that lets it switch between deep thinking for complex problems and direct responses for simple queries.
What it can do
GLM-4.5 performs strongly in benchmarks related to agents, reasoning, and coding. It scored 70.1 per cent on TAU-Bench, which tests agent capabilities, and 91.0 per cent on AIME 24, a mathematics reasoning test. On SWE-bench Verified, which evaluates software engineering skills, it achieved 64.2 per cent. These results place it third among all evaluated models, including both proprietary and open-source systems. It ranks second specifically on agentic benchmarks. The Air version also shows competitive results with a score of 59.8 while maintaining higher efficiency.
How to use it
Zhipu AI released the model weights under the MIT licence, permitting commercial use and secondary development. Users can download the base and hybrid reasoning models from Hugging Face or ModelScope. The company also provides API access through the Z.ai platform globally and the Zhipu AI Open Platform in mainland China. Code for running the model is available for popular frameworks like transformers, vLLM, and SGLang. FP8 versions are provided to reduce memory requirements for inference on supported hardware.
Questions people ask about GLM-4.5
What is the difference between GLM-4.5 and GLM-4.5-Air?
GLM-4.5 has 355 billion total parameters with 32 billion active, while GLM-4.5-Air is more compact with 106 billion total parameters and 12 billion active. Both share the same architecture and capabilities, but the Air version is designed for higher efficiency and lower resource usage.
Can I use GLM-4.5 for commercial projects?
Yes, the model is released under the MIT open-source licence. This permits commercial use, modification, and distribution. You can download the weights from Hugging Face or ModelScope and integrate them into your own applications without paying licensing fees.
How does the hybrid reasoning mode work?
The model offers two modes. Thinking mode engages deeper processing for complex reasoning and tool usage tasks. Non-thinking mode provides immediate responses for simpler queries. Users can control this behaviour via API parameters when sending requests to the model.
Good reading and sources
There is no English Wikipedia article on GLM-4.5 yet, so the facts on this page come from the sources listed above: huggingface.co/zai-org/GLM-4.5, arxiv.org/abs/2508.06471. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.