Qwen (3.5)

What it is

Qwen is Alibaba Cloud's family of open-weight models, released across an unusually wide span of sizes — from small models runnable on a phone or laptop up to large models competitive with other frontier-scale offerings. That range has made Qwen one of the most widely fine-tuned and self-hosted open model families alongside Llama, particularly popular for teams that want a single family covering everything from lightweight edge deployment to serious general-purpose use.

Where these numbers come from

Qwen 3.5, released February 2026, is a native vision-language model with 397 billion total parameters (17 billion active per pass) — a Mixture of Experts design like several other current-generation models. Alibaba has also previewed Qwen 3.6 Plus with a 1M-token context window via API, though not yet released as open weights. The MMLU figure above reflects the prior Qwen2.5 generation; update this entry via the admin dashboard once official Qwen 3.5 benchmarks are published.

Notable traits

• Released across a wide range of parameter sizes in the same generation, from roughly half a billion up to hundreds of billions of parameters.
• Includes dedicated variants specialized for coding (Qwen-Coder) and mathematics (Qwen-Math) alongside the general-purpose line.
• Fully open weights under a license permitting most commercial use, contributing to broad adoption outside China as well as within it.