Alibaba releases Qwen 3.5: open-weight models from 7B to 235B MoE with 128K context

Imagine having a highly capable AI assistant that speaks Italian, Chinese, English, and dozens of other languages fluently — and you can run it entirely on your own computer without paying any subscription. That is essentially what Alibaba just released with Qwen 3.5.

The family comes in several sizes. The smallest model has 7 billion parameters and can run on a mid-range gaming PC. The largest is a special 235-billion-parameter model that uses a technique called Mixture of Experts — think of it as assembling a team of specialists rather than relying on a single generalist, which makes it much cheaper to run than its size suggests.

One standout feature is the 128,000-token context window across all variants. That means the model can read and reason over roughly 200 pages of text in a single pass — useful for analyzing long contracts, research papers, or entire codebases without splitting documents into chunks.

The licensing is arguably the biggest news: Apache 2.0 means anyone — individual developer or large enterprise — can download, run, modify, and commercialize these models for free. Tools like Ollama make local deployment as simple as a single command.

Independent benchmarks show that Qwen 3.5-72B matches or beats expensive commercial models like GPT-5.5 on several tasks, particularly multilingual understanding and reasoning. For organizations looking to cut AI costs, keep data on-premises, or avoid vendor lock-in, Qwen 3.5 is one of the most compelling open-weight releases of 2026.