Zhipu AI Open-Sources GLM-4.1V-9B-Thinking with Chain-of-Thought Reasoning

Zhipu AI's GLM team has released a new open-source vision-language model on Hugging Face, the GLM-4.1V-9B-Thinking, marking a significant step forward in bringing reasoning capabilities to multimodal AI. Built on the GLM-4-9B-0414 foundation model, it is the first in the series to incorporate a chain-of-thought thinking paradigm. The model has already garnered over 440,000 downloads and nearly 800 likes since its release on July 22.

Unlike traditional vision-language models that directly produce answers from visual input, GLM-4.1V-9B-Thinking generates intermediate reasoning steps before arriving at its final response. This approach, powered by reinforcement learning, significantly improves accuracy, completeness, and interpretability in complex visual reasoning tasks such as mathematical problem-solving and multi-step analysis.

The model's benchmark performance is particularly striking. Among 28 standard evaluations, it achieved the best results among all 10B-parameter-class models on 23 tasks — and on 18 of those, it surpassed Qwen-2.5-VL-72B, a model with eight times as many parameters. This demonstrates that reasoning enhancement can help smaller models close the gap with, or even exceed, much larger architectures.

Technical specifications include support for 64K-token context windows, arbitrary aspect ratio image inputs, and resolutions up to 4K. The model is bilingual, supporting both Chinese and English. Zhipu AI also open-sourced the base model GLM-4.1V-9B-Base alongside the thinking variant, encouraging further research into the frontiers of VLM capabilities.

Deployment is straightforward through the Transformers library (version 4.57.1+), with additional support for vLLM and SGLang inference frameworks. The model card provides ready-to-use code for both single-image inference and server deployment, lowering the barrier for developers to experiment with reasoning-enhanced multimodal AI.

The release signals a shift in the open-source VLM landscape toward reasoning-first architectures. While reasoning improvements have largely been confined to pure language models in the past year, GLM-4.1V-9B-Thinking brings these capabilities into the multimodal domain, opening new possibilities for complex problem-solving, long-context understanding, and multimodal agent applications.

Looking ahead, the model raises practical questions for the community: how to balance reasoning depth with inference speed on limited hardware, and whether the thinking paradigm will become a standard component of future vision-language model designs. These are likely to be active areas of discussion in the coming weeks.