Launches on 2026-09-20

A 600B mixture-of-experts flagship with 27B active per token and a million-token window, at $1.00 per million in and $2.70 out — priced well under the frontier it is measured against.

A 7B image model that generates and edits in one checkpoint, with native transparent output — and, unlike every Qwen image release before it, a licence that forbids commercial use.

An open-source kernel for service firms that sell work: an agent drafts the deliverable in a sandbox with no credentials, and nothing reaches a client until a person approves it.

An online store run by chatting with three AI assistants for the storefront, marketing and logistics — free to start, with the card fee on the free plan roughly double the paid rate.

Go-to-market workflows for developer-tool companies: find developers already hitting the problem you solve, audit how AI search describes your product, and track comparable tools.

Alibaba's full-duplex speech-to-speech model on Model Studio, released 20 September 2026, with a 262,144-token context, function calling, web search and voice cloning; audio output costs $24 per million tokens in Singapore.

A 27B open-weight fine-tune of Qwen3.8-27B for emails, messages and short stories, with a 262,144-token context; weights under CC BY-NC 4.0 since 22 September, hosted API $0.24 in and $0.90 out per million tokens.

Jared Palmer's open-source family of small decision models (0.8B, 4B and 9B) on Qwen3.5 that return typed yes/no, choice and score probabilities, released 20 September 2026 under Apache 2.0 with training code and evaluation data.