Skywork-R1V3: A 38B Open Multimodal Model That Puts Reinforcement Learning First
Skywork-R1V is an advanced multimodal AI model series developed by Skywork AI, specializing in vision-language reasoning.
At a glance
- What is it?
- Skywork-R1V3 is a 38B parameter vision-language model that uses reinforcement finetuning to push multimodal reasoning scores. This review covers what it does, how to run it, and where its claims need scrutiny.
- Who is it for?
- Adopt Skywork-R1V3 if you need a permissively licensed (MIT) 38B multimodal model for reasoning-heavy tasks like math or logic from images, and you have the GPU memory for inference. Skip it if you need a small model or want a fully documented training pipeline, since the repository only provides weights and inference code.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 49 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What Skywork-R1V3 Solves and Who It Targets
Skywork-R1V3 is a 38 billion parameter multimodal model that answers questions requiring both visual understanding and multi-step reasoning. The README positions it as a state-of-the-art open-source model for tasks like math problems from diagrams, physics questions with figures, and logic puzzles that need visual clues. The target audience is researchers and engineers who need a model that can reason about images, not just describe or caption them. The project's own description emphasizes reinforcement finetuning, which means the model was post-trained with reinforcement learning rather than only supervised fine-tuning. This is a deliberate choice for improving reasoning ability, as the team reports significant gains on benchmarks like MMMU, where they claim 76.0 accuracy.
How the Model Is Built and What That Means
The README states that Skywork-R1V3 uses InternVL3-38B as the base model and applies reinforcement learning in post-training. The technical report (linked as a PDF) presumably details the RL algorithm, but the README only mentions that the model is trained via RL finetuning. The model family started with Skywork-R1V in March 2025, which the team called the first industry open-sourced multimodal reasoning model with visual chain-of-thought. Later iterations, R1V2 and R1V3, improved performance. The architecture is not described in the README beyond the base model, so you must treat the model as a black box: it takes an image and a question, and produces a reasoning trace and an answer. The key mechanism is that RL encourages the model to produce chain-of-thought reasoning before answering, which is why it excels at benchmarks that require multi-step logic. This is a different approach from models that only use supervised fine-tuning on reasoning traces.
Getting It Running: Commands and Configuration
The repository provides a straightforward path to local inference. First, clone the repo and navigate to the inference directory: git clone https://github.com/SkyworkAI/Skywork-R1V.git and then cd skywork-r1v/inference. For Transformers, create a conda environment with Python 3.10 and run bash setup.sh. For vLLM or evaluation, create another environment and run bash ./eval/vlmevalkit/build_env.sh. Then run inference with a script. The Transformers command is: CUDA_VISIBLE_DEVICES="0,1" python inference_with_transformers.py --model_path path --image_paths image1_path --question "your question". The vLLM command is similar but includes --tensor_parallel_size 4, which suggests you need multiple GPUs. The model weights are hosted on Hugging Face, so you must download them separately. The setup scripts handle dependencies, but there is no mention of a requirements.txt in the README, so the setup.sh is your only guide.
Benchmark Claims and How to Read Them
The README includes a large table comparing Skywork-R1V3 against QVQ-72B, InternVL-78B, Qwen-72B, Claude 3.7, and GPT-4o across many benchmarks. Skywork-R1V3 wins on several, including MMMU (76.0), MMMU-pro (55.4), MathVista (77.1), and HallusionBench (61.3). Some cells have an asterisk indicating results from the authors' own evaluation framework, which is a red flag: those numbers are not independently verified. The table also shows that Skywork-R1V3 does not win everywhere. Claude 3.7 beats it on EMMA, MathVision, SeePhys, and LogicVista, and Qwen-72B beats it on MMBench-en-1.1. So the claim of state-of-the-art is selective. The model is strong on math and logic but weaker on some general benchmarks. You should not treat these numbers as gospel; they are self-reported and the evaluation code is provided, but running it yourself is the only way to confirm.
A Real Limitation: Model Size and Hardware Demands
Skywork-R1V3 is a 38B parameter model, which is not small. The README mentions an AWQ quantized version of R1V2 that supports single-card inference on GPUs with more than 30GB memory, but for R1V3, no such quantized version is mentioned. The vLLM inference command uses --tensor_parallel_size 4, which implies you need at least four GPUs to run the model comfortably. If you have a single consumer GPU, you cannot run this model directly. The Transformers script also requires CUDA_VISIBLE_DEVICES with multiple device IDs. This is a significant barrier for individual developers or small teams without access to multi-GPU servers. The model is also not designed for edge or mobile deployment. If you need a lightweight model, this is the wrong tool.
Maintenance and Upgrade Cost
The repository has no releases listed, and the last push is dated July 2026, which is after the R1V3 release in July 2025. That suggests active development, but the README does not describe a roadmap or a versioning scheme. The project provides inference code and an evaluation directory, but no training code or data, so you cannot fine-tune the model further without external resources. The maintenance cost is low for users: you clone the repo, set up the environment, and run inference. However, if you want to reproduce the benchmark results, you must download the vlmevalkit and run the evaluation scripts, which can be time-consuming. The license is MIT for the code, and the base model InternVL3-38B is also MIT, which permits commercial use and modification. This is a permissive license, but you should verify the exact terms of the base model's license on its Hugging Face page, because the README only states it is MIT.
Alternatives and How They Differ
The most direct alternative is Qwen2.5-VL, which is mentioned in the comparison table. Qwen2.5-VL (72B) is a larger model that beats Skywork-R1V3 on MMBench-en-1.1 and MMstar. The key difference is that Qwen2.5-VL is trained with a different approach, likely a mix of supervised fine-tuning and RL, but it is not specifically optimized for chain-of-thought reasoning. Another alternative is InternVL3-78B, which is the larger sibling of the base model. InternVL3-78B beats Skywork-R1V3 on MMstar as well. The difference in approach is that Skywork-R1V3 applies RL finetuning on top of InternVL3, which is a base model that may already have some reasoning ability. If you need a model that is better on general visual question answering, Qwen2.5-VL might be a better choice. If you need a model that excels on math and logic, Skywork-R1V3 has the edge according to the reported numbers.
Editorial conclusion
Adopt Skywork-R1V3 if you need a permissively licensed (MIT) 38B multimodal model for reasoning-heavy tasks like math or logic from images, and you have the GPU memory for inference. Skip it if you need a small model or want a fully documented training pipeline, since the repository only provides weights and inference code. Before adopting, verify the model's performance on your specific benchmark, because the reported scores come from the authors' own evaluation and may not generalize. Also check the base model's license terms, even though InternVL3-38B is MIT, to ensure your use case is covered.
Community notes