Qwen3.5 (github.com)

Qwen3.5 has been officially released, showcasing a significant advancement in foundation models with a focus on multimodal learning and enhanced performance across various benchmarks. This model introduces a unified vision-language foundation that exhibits superior cross-generational capabilities compared to its predecessor, Qwen3, as well as improved architectural efficiency through a hybrid design utilizing Gated Delta Networks and sparse Mixture-of-Experts. Additionally, Qwen3.5 scales reinforcement learning across millions of agents, enhancing its adaptability for real-world applications, and expands linguistic support to 201 languages, ensuring global accessibility. This release is pivotal for the AI/ML community as it not only enhances the capabilities of language models but also emphasizes efficiency in training and deployment. With near-100% multimodal training efficiency and reduced latency costs, developers and enterprises can leverage Qwen3.5 for various applications, including coding assistance and visual understanding tasks. The model's infrastructure supports a wide array of frameworks, making it easier for users to integrate it into their own applications via the Qwen API and available tools like Qwen Code and Qwen Agent. Overall, Qwen3.5 is positioned to redefine the landscape of AI application development with its innovative features and robust support mechanisms.