Skywork-R1V is an open-source multimodal reasoning model series from Skywork AI, with RL finetuning and inference scripts for Transformers and vLLM. The repo includes setup and run instructions, plus evaluation benchmarks and licenses under MIT.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Skywork-R1V3 is an open-source multimodal reasoning model family by Skywork AI. It includes model weights and inference code, with reinforcement learning finetuning mentioned in the README.
How it works
The repository provides inference scripts for two pathways: using Transformers and using vLLM. It includes environment setup steps and commands to run inference with image inputs and questions. It references a base model InternVL3-38B and includes evaluation benchmarks in its reports.
Getting started
git clone https://github.com/SkyworkAI/Skywork-R1V.git
cd skywork-r1v/inference
# For Transformers
conda create -n r1-v python=3.10 && conda activate r1-v
bash setup.sh
# For vLLM/evaluation
conda create -n r1v-vllm python=3.10 && conda activate r1v-vllm
bash ./eval/vlmevalkit/build_env.sh
# Using Transformers
CUDA_VISIBLE_DEVICES="0,1" python inference_with_transformers.py \
--model_path path \
--image_paths image1_path \
--question "your question"
# Using vLLM
python inference_with_vllm.py \
--model_path path \
--image_paths image1_path image2_path \
--question "your question" \
--tensor_parallel_size 4
Getting started
The README includes commands to clone, set up environments for two runtimes, and run inference scripts as shown above. It also notes license terms under MIT and credits a base model from InternVL3-38B.
Recent releases
The latest releases section lists none.
Traction
3162 stars, 281 forks, 35 open issues as of the repository metadata.
Behind the repo
The project is developed by Skywork AI and references the InternVL3-38B base model from OpenGVLab for licensing.
Caveats
Licensed under MIT; commercial use permitted; modification and distribution allowed. The README notes no explicit age caveats beyond creation date 2025-03-15 and last push 2026-07-29.