Skywork-R1V: open weights and a closed API share one repo

What it does

Skywork-R1V is a family of vision-language models built to reason over images using chain-of-thought and reinforcement learning. The repository hosts downloadable weights and inference code for the earlier 38-billion-parameter models (R1V through R1V3), alongside Python utilities for batch-testing the latest R1V4-Lite, which is available only as a hosted API.

The interesting bit

The project maintains an unusual split personality: older checkpoints are fully open and quantized for consumer hardware, while the newest R1V4-Lite is a closed-source agent that writes and executes Python code and can call search tools. Benchmark tables in the README claim the 30-billion-parameter R1V4-Lite outperforms similarly sized Qwen3-VL variants and even some larger Gemini models on perception tasks like HIRbench and MME-Real.

Key highlights

• R1V3-38B is trained with RL in post-training and scores 76.0 on the MMMU benchmark.
• Quantized AWQ and GGUF releases target single-GPU inference (≥30 GB VRAM) or CPU-only setups.
• R1V4-Lite supports multi-turn reasoning with tool use, including optional web search for deep research.
• The repo provides a batch-testing suite with streaming, non-streaming, and planner modes for the API.
• All repository code is under the MIT license and permits commercial use.

Caveats

• R1V4-Lite is closed-source and API-only; local self-hosting is not possible.
• The README dedicates most of its space to the commercial API product, leaving the open-source model documentation comparatively sparse.
• R1V4-Lite is explicitly based on Qwen3-VL-30B-A3B-Instruct, but the README does not detail what training or architectural changes differentiate it from the base model.

Verdict

Grab the open weights if you need a locally runnable vision-reasoning model and have the VRAM. If you are hunting for a fully open, agentic multimodal system you can self-host, look elsewhere—the newest capabilities are locked behind Skywork’s API.