Getting Started with Qwen3.5 Vision-Language Models

Qwen3.5 is a new open-source vision-language model suite from Qwen Team, ranging from 0.8B to 397B parameters. It features a heterogeneous architecture that decouples parallelism across vision and language components, native FP8 training, Mixture-of-Experts routing, and an asynchronous RL framework achieving 3–5× training speed improvements. The tutorial walks through deploying Qwen3.5-122B on a DigitalOcean GPU Droplet using Ollama and Claude Code to generate a Python curling game, noting that while the model produces functional code, iterative refinement is slow on a single H200 GPU and results still require significant manual improvement.