GLM-4.1V-Thinking
Open-source vision-language model with reasoning focus
GLM-4.1V-Thinking is a 9-billion parameter vision-language model released by the GLM-V Team in 2025. It uses reinforcement learning to improve reasoning in complex tasks like mathematics and video understanding. The model supports bilingual Chinese and English usage and handles high-resolution images.
- Uses reinforcement learning for enhanced reasoning
- Supports 64k context length and 4K images
- Outperforms larger models on many benchmarks
- Open-source weights available for research
- Bilingual support for Chinese and English
What it is
GLM-4.1V-Thinking is a vision-language model designed to handle complex multimodal tasks. Developed by the GLM-V Team, it builds upon the GLM-4-9B-0414 foundation. The model focuses on enhancing reasoning capabilities through a specific training paradigm. It supports both Chinese and English languages. With 9 billion parameters, it aims to provide high performance in areas such as mathematics and long-context understanding. The system can process images up to 4K resolution and arbitrary aspect ratios.
How it came about
The model was introduced in July 2025 as part of a broader family including GLM-4.5V and GLM-4.6V. It represents an evolution from previous CogVLM2 and GLM-4V series models. The development team focused on creating a reasoning-centric training framework. This approach aimed to unlock the full potential of the vision foundation model. The release included open-source weights to support further research into multimodal capabilities.
Inside the model
The architecture employs a Chain-of-Thought reasoning paradigm to improve answer accuracy and interpretability. Training utilizes Reinforcement Learning with Curriculum Sampling (RLCS). This method enhances capabilities across diverse tasks including STEM problem solving and GUI-based agents. The model supports a context length of 64,000 tokens. It processes visual inputs with support for arbitrary aspect ratios and resolutions up to 4K. The base model GLM-4.1V-9B-Base was also released to aid research.
What it can do
The model demonstrates strong performance in mathematical reasoning and complex problem solving. It achieved top results among 10-billion parameter models on 23 out of 28 benchmark tasks. In some evaluations, it surpassed the much larger Qwen-2.5-VL-72B model on 18 tasks. Capabilities extend to video understanding, content recognition, coding, and long document interpretation. The system provides rich and interpretable answers compared to non-reasoning visual models.
How to use it
Weights are available openly for download via platforms like Hugging Face. The model can be run using the transformers library with specific version requirements. An API is provided through the Zhipu Foundation Model Open Platform. Code for inference, video reasoning, and web demo deployment is hosted on GitHub. This open approach allows developers and researchers to integrate the model into various applications.
Questions people ask about GLM-4.1V-Thinking
What makes GLM-4.1V-Thinking different from other vision models?
It incorporates a thinking paradigm and reinforcement learning to improve reasoning. This allows it to handle complex tasks like mathematics and long-context understanding better than traditional non-reasoning visual models.
Can this model process video inputs?
Yes, the model supports video reasoning. Code examples for video processing are available on the project GitHub repository, allowing users to deploy it for video understanding tasks.
How large is the context window?
The model supports a context length of 64,000 tokens. This enables it to handle long documents and complex multi-step reasoning tasks that require maintaining information over extended sequences.
Good reading and sources
There is no English Wikipedia article on GLM-4.1V-Thinking yet, so the facts on this page come from the sources listed above: huggingface.co/zai-org/GLM-4.1V-9B-Thinking, arxiv.org/abs/2507.01006. 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.