Alibaba’s Qwen 3.5 LLM now official

Alibaba has introduced Qwen 3.5, their most powerful multimodal large language model (LLM) yet. It’s positioned directly against today’s top closed models like GPT, Claude, and Gemini.

The flagship model packs 397 billion parameters, but only activates 17 billion at inference time. In other words, this makes it significantly more efficient than its raw size suggests.

Qwen 3.5 supports a one‑million‑token context window. For reference, this is big enough to handle more than 700,000 words, a medium-sized codebase, or over an hour of video in a single prompt.

Benchmarks highlight strong performance in instruction following, graduate-level science questions, agentic abilities, document understanding, and video reasoning. Because it is fully multimodal, Qwen 3.5 can accept images and video as input, then perform tasks like question answering, document analysis, and more.

It also offers presentation generation and even coding full 3D racing games or web front-ends from scratch. The model also demonstrates robust spatial reasoning, solving puzzles like mazes and Sudoku from images.

In addition, the LLM can be wired into messaging apps like WhatsApp and Telegram via OpenAI-compatible tooling. Alibaba has open‑sourced Qwen 3.5, with detailed instructions for local deployment via its GitHub repository.

The full model is around 87 GB, so it targets high-end GPUs and data-center setups rather than typical consumer hardware. However, users can also access it for free through the Qwen Chat platform by selecting it via the model dropdown.