Qwen3.5 Delivers Open-Source Vision-Language Advances and Deployability

Qwen Team releases Qwen3.5, an open-source vision-language model suite ranging from 0.8B to 397B parameters, including activated 17B and 122B variants. The models use heterogeneous parallelism, sparse activations, and an FP8 training pipeline plus an asynchronous RL framework to boost throughput, reduce memory, and scale multimodal training. The article includes deployment guidance using Ollama and GPU droplets and a code-generation demo.