Practical overview
When it fits
Use Revisual R1 when you need a documented research codebase and model checkpoint for local image-and-text reasoning, or want to examine a staged multimodal training approach.
Who it is for
- Researchers exploring multimodal reasoning and staged reinforcement learning.
- Developers building image-and-text reasoning prototypes or visual question-answering systems.
- Teams evaluating multimodal models on mathematical reasoning datasets.
Documented uses
- Generate answers to questions about images using a local model checkpoint.
- Evaluate a model on supported multimodal mathematics datasets.
- Study or adapt the documented cold-start and staged reinforcement-learning workflow.
Requirements & limits
The repository is public and identifies Apache-2.0 as its license. The README also displays an MIT license badge; check the repository’s license file for the applicable terms.
- The model is a 7B checkpoint based on Qwen2.5-VL-7B-Instruct.
- The repository documents Python 3.10 and Conda setup, with an example flash-attn 2.7.3 installation option.
- The evaluation script supports named datasets including Geo3K, WeMath, MathVista, MathVerse and MathVision.
- The documented evaluation workflow requires an OpenAI API key; a separate example shows local serving with lmdeploy.
- The README includes inference, cold-start training, staged reinforcement optimization and checkpoint-merging instructions.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the Revisual R1 Python environment
Create the documented Python environment and install the repository in editable mode.
Before you start
- Conda
- Python 3.10
- A local checkout of the Revisual-R1 repository
Steps
Create and activate a Conda environment named revisual with Python 3.10.
From the repository directory, run pip3 install -e .
If flash-attn installation is an issue, the README gives a flash-attn 2.7.3 wheel as an example installation option.
Expected outcome
The repository dependencies are installed in the Conda environment; this setup step alone does not run inference or training.
Guide limitations
- The README presents the flash-attn wheel as an example, not a guarantee that it suits every environment.
- No hardware requirements are specified in the documented setup steps.
Run local inference with an image and question
Use the documented inference script with a local model checkpoint and image input.
Before you start
- A ReVisual-R1 model checkpoint
- An image file
- The repository’s documented Python environment
Steps
Set MODEL_PATH to the local model checkpoint and provide an image path and question.
Run infer.py with the model path, image path, question, prompt and generation options, or modify and run inference/inference.sh.
Expected outcome
The documented command requests a model response to an image-based question.
Guide limitations
- The README’s example uses placeholder values such as xxx for the question and does not specify hardware requirements.
- The command assumes a prepared local checkpoint and environment.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- CSfufu/Revisual-R1 · Official homepage title · retained in its source language