Qwen3.5 aims to position Alibaba alongside GPT and Claude

"Qwen3.5 is available via Hugging Face and is released under an open-source license. With this, Alibaba is explicitly targeting developers and research institutions that want to work with the model themselves. The system can process very long prompts, up to 260,000 tokens, and can be scaled further with additional optimizations. This makes it suitable for complex applications such as extensive document analysis and code generation."

"The architecture of Qwen3.5 is based on the so-called mixture-of-experts principle. Instead of one large neural network, the model uses multiple specialized networks, only a limited number of which are active per task. This significantly reduces the required computing power without compromising performance. Although the total model has nearly 400 billion parameters, only a fraction of these are used per prompt."

"An important part of this is the way the model handles attention mechanisms. Whereas traditional attention mechanisms quickly consume a lot of memory with long inputs, Qwen3.5 combines classic approaches with a lighter variant that requires less memory. This makes the model more scalable for applications with large amounts of context. Qwen3.5 also uses a so-called gated delta network. This technique helps the model temporarily discard irrelevant information, making the learning process during training more efficient."

Qwen3.5 is an open-source large language model released on Hugging Face under an open license and aimed at developers and research institutions. The model supports extremely long contexts up to 260,000 tokens and can scale further with optimizations, making it suitable for extensive document analysis and code generation. It handles more than 210 languages and dialects and accepts image inputs including graphs. The architecture uses a mixture-of-experts design with nearly 400 billion total parameters while activating only a subset per task to reduce compute. Attention mechanisms are blended with a lighter variant to save memory, and a gated delta network helps discard irrelevant information during training to improve efficiency. Performance comparisons place it competitive with GPT-5.2 and Claude 4.5 Opus in several areas.

Read at Techzine Global