Alibaba's change of course: Qwen-Image-2.1 and the thin line between open weights and open source
On 20 September 2026 Alibaba's development team published Qwen-Image-2.1 on the GitHub, Hugging Face and ModelScope repositories. The visual generation component has 7 billion parameters with 32 single-stream DiT layers, alongside the Qwen3-VL text encoder with 8 billion parameters, a flow matching scheduler and a 64-channel RGBA VAE with 16× compression. The architecture supports native generation with an alpha channel for transparency, editing guided by masks or annotations, and identity preservation. Comfy.org's integration documentation reports a native resolution of 2048×2048 pixels and an attribution requirement with the wording "Built with Qwen". Some operational inconsistencies remain: the official repository caps reference images for multi-subject composition at 10, while Comfy.org's integration documentation indicates a threshold of up to 16.
The most significant change lies in the legal framework. Compared with the first 20-billion-parameter iteration from August 2025 and the 2512 update from December 2025, both distributed under the permissive Apache 2.0 licence, version 2.1 adopts the "Qwen RESEARCH LICENSE AGREEMENT". The text, which names Hangzhou Tongyi Laboratory Technology Co., Ltd. as licensor, grants use "FOR NON-COMMERCIAL PURPOSES ONLY", defining the non-commercial scope solely as research or evaluation. Anyone intending to use the model commercially has to negotiate a separate agreement. The licence of the intermediate Qwen-Image-2.0 (February 2026) cannot be verified — its Hugging Face card is not accessible — so there is no way to establish whether the restriction arrives with 2.1 or was already in force.
On performance, The Decoder (20 September 2026) writes that the model beats most closed models on Qwen's own benchmark, noting that independent benchmarks are still awaited. Neither the Hugging Face model card nor the GitHub repository, however, includes comparison tables or official numerical scores, and no third-party verification is available yet. The absence of public metrics, together with the fact that no statement explaining the licence choice was found on official channels, leaves the presumed performance superiority and the strategic reasons for the restriction within the perimeter of unverified claims.
Distributing a model's weights while keeping a commercial-use restriction is a legitimate choice, but it marks out a precise perimeter: technological openness without a free licence remains a revocable concession, not an open standard for the ecosystem. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I opened the LICENSE file in the official QwenLM/Qwen-Image-2.1 repository on GitHub and took from it the non-commercial clause, the definition of "Non-Commercial", the release date (20 September 2026) and the licensor's name. The official Hugging Face model card confirms the same licence label and the model specifications (7B, RGBA, editing). For comparison I opened the model cards of Qwen/Qwen-Image (20B) and Qwen/Qwen-Image-2512 (20B): both Apache 2.0. The Qwen-Image-2.0 page returned 401, so that detail stays out. Independent confirmations: The Decoder of 20 September 2026 (commercial restriction, in-house benchmarks only) and Comfy.org's technical documentation (specifications, native resolution, "Built with Qwen" attribution). Where the sources disagree — the number of reference images — I put it among the uncertainties instead of picking one at random.
- Incertezze
- The licence of the intermediate Qwen-Image-2.0 (February 2026) is unverified: its Hugging Face page is not accessible, so we cannot tell whether the restriction arrives with 2.1 or was already in 2.0. Reference images are 10 in the official repository and 16 in Comfy.org's documentation. There are no independent benchmarks and no official scores: the claimed superiority over closed models remains the vendor's word. There is no post dedicated to 2.1 on the qwen.ai blog, and the Qwen team has not publicly explained the licence choice.
- Perché pubblicarla
- The verifiable fact is not the model, it is the licence: a family born under Apache 2.0 releases a version usable only for research and evaluation. It concerns anyone in Italy building products on open Chinese weights — the update is not as free as it looks — and it lets us show, documents in hand, the difference between "downloadable weights" and "open source". And without independent benchmarks, the release deserves telling without amplifying the vendor's numbers.