OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Visualthinker R1 Zero

Visualthinker R1 Zero is a Python research repository for reproducing an R1-Zero-style visual-reasoning approach with Qwen-VL-2B. It documents SAT dataset preparation, GRPO reinforcement-learning training, an SFT comparison path and CVBench evaluation scripts. The repository also links to a released model checkpoint.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Visualthinker R1 Zero — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when you need a documented research workflow for training and evaluating vision-centric reasoning models, rather than a ready-made image-questioning service.

Who it is for

  • Researchers studying reinforcement learning for multimodal reasoning
  • Developers reproducing the repository's SAT training and CVBench evaluation workflow

Documented uses

  • Prepare the SAT dataset for the documented training workflow
  • Launch GRPO training for a 2B visual-reasoning model
  • Train an SFT comparison model
  • Evaluate model variants on CVBench with or without a reasoning prompt
Before you choose

Requirements & limits

  • The README estimates 4×80 GB for GRPO full fine-tuning in AMP on the 2B model.
  • The repository documents both reasoning-prompt and no-reasoning-prompt evaluation options.
  • A model checkpoint is linked from the repository, but detailed inference instructions are not included in the README.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare the SAT dataset

Run the repository's documented SAT dataset preparation step.

Before you start

  • A local copy of the repository
  • A shell environment to run the provided scripts

Steps

  1. Run the command below from the repository root.

    Documented command
    bash setup.sh

  2. Change to src/data/SAT.

  3. Run the command below.

    Documented command
    bash prepare_dataset.sh

Expected outcome

The documented setup and SAT dataset preparation scripts have been run.

Guide limitations

  • The README does not specify dataset contents, download requirements or how to verify preparation.
Documentation 1 ↗
Source-based practical guide

Train the GRPO visual-reasoning model

Start the repository's documented GRPO training path on SAT.

Before you start

  • A prepared SAT dataset
  • The repository setup completed
  • Estimated hardware requirement: 4×80 GB for GRPO full fine-tuning in AMP on the 2B model

Steps

  1. Change to src/open-r1-multimodal.

  2. Run the command below.

    Documented command
    bash run_grpo_SAT.sh

  3. Adjust open-r1-multimodal/configs/zero3.yaml or zero2.yaml as indicated by the README.

Expected outcome

The documented GRPO training script is launched for the SAT dataset.

Guide limitations

  • The hardware figure is marked as estimated.
  • The README does not provide detailed configuration guidance or describe expected training duration.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
18 October 2026
Original website title
turningpoint-ai/VisualThinker-R1-Zero · Official homepage title · retained in its source language
How we review sourcesSuggest a correction