VisualThinker-R1-Zero is a research-oriented open-source project focused on multimodal visual reasoning, with the goal of reproducing the training paradigm of DeepSeek-R1-Zero in visual reasoning tasks.
In terms of purpose, this project is dedicated to exploring and verifying whether, in visual reasoning tasks, it is also possible, as with pure text reasoning, to enable a model to spontaneously develop an "aha moment" and longer reasoning responses through pure reinforcement learning (without SFT and without a reward model). The research team successfully observed this phenomenon for the first time on a non-SFT multimodal model with only 2B parameters.
In terms of main capabilities, the project is based on the Qwen-VL-2B model and uses the GRPO reinforcement learning algorithm for training on the SAT dataset. The research shows that vision-centric tasks can also benefit from enhanced reasoning ability; during training, self-reflection behavior similar to DeepSeek R1 was also observed, with the model able to rethink and correct its own mistakes. The project provides a complete record of training dynamics and has open-sourced model checkpoints (released on Hugging Face).
In terms of usage, the project repository provides the specific code commands needed to reproduce training and evaluation: including training a non-SFT model with GRPO on the SAT dataset to reproduce the multimodal aha moment; training a non-SFT model to obtain an SFT model for comparison; and commands for reproducing evaluation results on CVBench, covering the Base + GRPO (VisualThinker R1 Zero) model, the Base model, the Instruct + GRPO model, and the Instruct model respectively. The complete experiment logs are marked as "coming soon."
In terms of applicable scenarios, this project is mainly intended for researchers and developers engaged in multimodal visual reasoning, reinforcement learning training, and research on emergent model behavior, and can be used to reproduce experiments, compare different training strategies, and conduct further research on this basis.
In terms of limitations, the project materials only describe its research methods, training commands, and evaluation process, and do not provide specific information about pricing, commercial use authorization, deployment platforms, or performance metrics; in addition, the complete experiment logs and some other information are not yet available.
