Skywork-R1V
Skywork AI (Kunlun Tech) · Other
Skywork-R1V is an open multimodal reasoning model series from Skywork AI. It reads images and works through multi-step chains of thought to solve math, science and chart problems.
About Skywork-R1V
What Is Skywork-R1V
Skywork-R1V is a family of vision language models built by Skywork AI, the research arm of Kunlun Tech. As an open source multimodal AI release, the series puts the weights in your hands. The idea behind the series is simple to state and hard to pull off: take a model that already reasons well in text and give it the ability to reason over images without losing that skill. Skywork calls this moving strong reasoning from text to vision. Where most multimodal models recognize what's in a picture, R1V works through the steps behind it.
The first R1V shipped in March 2025 as an open weight release, and the team has iterated fast since. R1V2 followed in April, R1V3-38B arrived in July with 76.0 on the MMMU benchmark, and R1V4-Lite landed in November as a lighter agentic model. Each step moved the series from pure reasoning toward models that act. You can download the weights from Hugging Face, run them locally, or call the hosted API.
Two limits matter up front. The 38B models are large enough that running them yourself needs serious hardware unless you pick a quantized build. And R1V is a research-grade release, so the tooling around it assumes you know your way around Python and a model server.
Getting Started
- Pick a version from the Hugging Face collection that matches your hardware, from R1V3-38B down to a quantized build
- Clone the Skywork-R1V GitHub repo and set up the Python environment with the provided script
- Load a model with Transformers or vLLM and point it at an image plus a question
- For a hosted route, create an account on the Skywork API platform and call the completions endpoint with your image and prompt
Product Information
A quick look at Skywork-R1V's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- AI researchers
- Machine learning engineers
- Students and self-learners
- Startups on a budget
Tasks
- Visual math and science problems
- Chart and diagram analysis
- Document understanding
- Medical and technical image review
- Agentic image tasks
Scenarios
- A student photographs a handwritten problem
- A researcher compares benchmarks
- A small team needs vision without a big bill
- A developer reads a dense chart
- A support team sorts scanned forms
Key features
Multimodal Chain-of-Thought Reasoning
The core of Skywork-R1V is that it reasons in steps rather than pattern-matching an answer. Give it a geometry problem with a diagram and it works through the logic before it lands on a result. The team built the series around transferring text reasoning into vision, so the model keeps its math and logic skills while it reads an image. That combination is what separates it from a plain image captioning model.
Adaptive Chain-of-Thought Length
Not every question needs a long train of thought. R1V adjusts how much reasoning it does based on how hard the problem looks and how clear the image is. A simple recognition query gets a short path, while a multi-step science problem gets the full treatment. This keeps inference efficient and cuts the overthinking that slows other reasoning models down.
State-of-the-Art Open Benchmarks
According to Skywork, R1V3-38B scores 76.0 on MMMU and 77.1 on MathVista, and the earlier R1V hit 94.0 on MATH500 and 72.0 on AIME. Those numbers put the open release in range with much larger proprietary models on cross-disciplinary tasks. The team publishes the evaluation code, so you can rerun the benchmarks rather than take the scores on faith.
Open Weights Under a Permissive License
The R1V3 model weights ship under the MIT license, which allows commercial use, modification and distribution. You can download them from Hugging Face, run them on your own GPU and build a product on top without paying a per-call fee. Few vision models of this class come with terms that open.
Quantized Builds for Smaller Hardware
Skywork publishes AWQ quantized versions of R1V and R1V2 that run on a single card with more than 30GB of memory. That drops the hardware bar from a multi-GPU rig to one workstation card, which matters if you want to test the model before committing to a cluster. The trade-off is a small hit to accuracy that shows up on the hardest benchmarks.
Active Image Operations in R1V4-Lite
The newest release, R1V4-Lite, does something the earlier models cannot. When a photo is blurry or the framing is off, it crops, zooms, rotates or locates a region on its own and then reasons over the better view. It can write a small tool to fix the image mid-task. The model is light enough to run in production and answers in a fraction of the time larger models take.
Pros and cons
Pros
- Open weights with an MIT license let you self-host and build on the model for free
- Strong benchmark results across MMMU, MathVista and MATH500 for a 38B model
- Reasoning in steps makes answers easier to check than a single-shot guess
- Quantized builds run on one GPU with 30GB of memory, cutting the hardware cost
- The hosted API is free to try on the Skywork platform
Cons
- Running the full 38B model yourself needs a serious GPU setup
- Quantized versions lose accuracy on the hardest tasks, so the trade-off is real
- Setup assumes Python and model-server skills, which puts it out of reach for non-technical users
- R1V4-Lite is billed as an agent, but its channel support is narrower than the reasoning core
Frequently asked questions
Skywork-R1V is an open series of multimodal reasoning models from Skywork AI. Each model reads an image and works through a chain of thought to answer a question, rather than just naming what it sees. The series spans R1V, R1V2, R1V3-38B and R1V4-Lite.
Related content
Explore related tools, skills, and articles for Skywork-R1V.
Skywork-R1V Alternatives
Odyssey 3
Odyssey · OtherOdyssey-3 is Odyssey's foundation world model. It simulates physics in real time, turns a prompt into an interactive environment, and adapts to robots, cars and games.
JarvisCore
Prescott Data · Coding · Productivity · OtherJarvisCore is an open source Python runtime where AI agents work as equal peers on an agent mesh, claim tasks from a shared ledger, and reach external services through a zero-trust broker.
Clef
Cloudflare · Business · OtherClef is Cloudflare's open-source decision model for Workers AI. It returns calibrated probabilities for typed questions, not generated text, so agents can classify, route, and act fast.