InternVL3.5 Drops Training Code and CascadeRL: What You Need to Know

Multimodal models are getting more capable by the day, but the real bottleneck has always been how to train them effectively. That's why it's exciting to see InternVL3.5 open-source its full training pipeline and a new reinforcement learning approach called CascadeRL. If you've ever tried to fine-tune a vision-language model beyond the usual image captioning tasks, this is worth your time.

What It Does

InternVL3.5 is a family of multimodal models that combine vision and language understanding. The newly released code includes:

• The complete training setup for InternVL3.5 models (from scratch or fine-tuning)
• CascadeRL — a multi-stage reinforcement learning method designed for multimodal tasks
• Evaluation scripts and model weights

The repo lives at github.com/OpenGVLab/InternVL and has everything you'd expect: configs, data processing pipelines, and training logs.

Why It’s Cool

CascadeRL is the standout feature here. Most multimodal RL approaches apply a single reward signal across all modalities. CascadeRL instead breaks the training into stages where different reward models focus on different aspects — visual grounding, language coherence, reasoning chains — and then layers them. Think of it as curriculum learning meets RLHF, but designed for the messy reality of real-world vision-language tasks.

Other things worth noting:

• Reproducibility first. They provide exact configs and checkpoints, so you can verify results without guesswork.
• Memory efficient. The training code supports gradient checkpointing and mixed precision out of the box.
• Framework agnostic. It's built on standard PyTorch and DeepSpeed, so you can adapt it to your own setup.

How to Try It

If you want to get your hands dirty, the quickest path is:

git clone https://github.com/OpenGVLab/InternVL.git
cd InternVL
pip install -r requirements.txt

Then check out their quickstart notebook in the examples/ folder. For training your own model, the internvl_chat/train/ directory contains the main entry points and configs for both SFT and CascadeRL.

If you just want to run inference, they provide a simple script:

python demo.py --model-path OpenGVLab/InternVL3-5B

No need to train from scratch unless you want to customize the pipeline.

Final Thoughts

What I appreciate most about this release is that it doesn't just dump a model and walk away. By open-sourcing the training infrastructure and CascadeRL, they've given developers a real sandbox to experiment with multimodal RL strategies. Whether you're building a visual QA system, an agent that needs to "look" at things, or just curious about how RL scales to multi-modal tasks, this repo is a solid foundation. It's not hype — it's a practical tool that actually works.

Find more open source projects and tools at @githubprojects