Enhancing NVIDIA Nemotron 3.5 Super 120B VL with agentic web development capabilities

We recently received early access to NVIDIA's Nemotron 3.5 Super 120B VL and had the opportunity to run some experiments on the new model. Our goal was to extend it with agentic web development capabilities via reinforcement learning. Using an ACP-based harness with OpenCode v2 and a Playwright MCP server, the model can now visually evaluate its own work in a browser, just like a human developer would. Our RL-fine-tuned model delivers better results at lower token costs.

Normally, fine-tuning a 120B model in a generation-demanding RL setup is very expensive. However, our LoRA-based approach keeps resource requirements in check and makes a proof of concept like this feasible with a fairly small investment.

Once again, a big thank you to NVIDIA for the early access and their kind support throughout our experiments!

Check out our article on Hugging Face for more details.