Two open-weights releases this week change what a brand studio can run on its own machine. On 9 October Alibaba’s Qwen team released Qwen-Image-2.1-Turbo, a 7B image model that generates and edits at 2K in eight denoising steps instead of forty. Three days earlier Kandinsky Lab put Kandinsky 6.0 on Hugging Face under MIT: five-second video with its own 44 kHz sound, lip-sync included. Both run on a gaming card. One of them you may not use commercially, and that is why it leads.
Qwen-Image-2.1-Turbo: a 2K image in eight steps, for research only
The Qwen repository’s changelog for 9 October lists two things: Qwen-Image-2.1-Turbo, “for image generation and editing in just 8 denoising steps”, and the Pro and Turbo APIs going live on Alibaba Cloud Model Studio (Qwen GitHub, 9 Oct). The model card puts Turbo on the same 7B architecture as the base model, defaults to the 2048 level with presets up to 2752×1536, takes reference images for editing and multi-reference composition, runs at CFG 1 with a saved eight-step schedule, and states: “This model is licensed under the Qwen Research License Agreement” (Hugging Face model card). Changing the step count does not override that schedule; only the saved one has been evaluated.
My read: The product tiles in One Group. Every Shelf. were generated with AI, eight categories of them, and the real cost of a wrong tile was the wait before I could see it was wrong. An eight-step model turns that wait into a conversation with the art director. That is what Turbo is for: iteration on a single workstation, where an unreleased pack never leaves the building. Two limits. The licence is research-only; a paid campaign built on these weights needs either a separate agreement or the API, which carries its own terms. And test it on your own failure cases first, which for me means Persian and Arabic typography on a pack. No demo set tells you that.
Kandinsky 6.0: open video with sound, and the render times that come with it
Kandinsky Lab’s readme dates the release to 6 October: Kandinsky 6.0 Video Lite (3B) and Pro (29B), both producing five-second clips with synchronized 44 kHz audio including lip-sync, from text or from an image, plus a separate super-resolution model that upscales to 1920×1080 (Kandinsky GitHub, 6 Oct). The weights carry an MIT licence; the Lite checkpoint downloads freely, while the Pro checkpoint is gated behind accepted conditions and contact details (Hugging Face, Pro · Lite). Native output is 480×864 at 24 fps. The readme’s own table gives one five-second Lite clip at 437 seconds on an RTX 4090 and a Pro HD clip at 1,194 seconds; on an H100 a Pro Full-HD clip takes 402 seconds.
My read: I built Mooo & Tiny, Break Time Is Snack Time as a 24-second music video: the track came first and every cut sat on it. Kandinsky generates picture and sound together, five seconds at a time, which is the opposite order and a real idea for jingles and mascot lines. Do the arithmetic before you promise it. A 24-second film is five clips; at Pro HD on a 4090 that is about a hundred minutes of render per take, before the upscale and before the take you reject. MIT is the most generous licence in this week’s drops. It is also a Russian lab’s model, so for some clients the question will be procurement, not quality.
GPT-6 reaches every ChatGPT user, and the answer becomes a UI
OpenAI rolled GPT-6 into the Chat tab for Plus, Pro, Business and Enterprise on 7 October, and to Free and Go the next day; paid tiers get GPT-6 Sol, free tiers GPT-6 Luna (OpenAI community, 7 Oct). The same release adds Intelligent UI: answers can now mix text with graphics, tappable buttons, forms, charts, maps, interactive diagrams, side-by-side comparisons, calculators and small in-chat games, with plain text kept where it works best. OpenAI’s wording: “Instead of fitting every question into the same conversational format, ChatGPT can combine UI, data, and actions” (OpenAI). It lands four days after OpenAI described visual ads inside ChatGPT, with first tests later this month (OpenAI, 5 Oct).
My read: Side-by-side comparisons is the phrase a brand team should read twice. Your pack is about to sit in a card next to a competitor’s, at phone-thumbnail size, chosen by a model, not by a planner. The test I used on the range reel applies: every pack had to read inside one tile of a grid on a phone, and the ones that failed were the ones with small type and low contrast. You do not control how ChatGPT renders you. You control the pack shot, the product name and the structured data it reads. Fix those this month, before the ads arrive.
What I’d do today
Put Qwen-Image-2.1-Turbo on one workstation, label it “test bench, research licence” in the folder name, and run your three hardest pack shots through it with the pack artwork as the reference. If it holds the typography, decide whether the API or a different model carries the paid work. Then ask ChatGPT to compare your top SKU with its nearest rival and look at the card. Need the production side? Here is how I work.
Sources
- Qwen (GitHub) — Qwen-Image-2.1 readme: 2026.10.09 Turbo release, Pro/Turbo APIs live; Qwen Research License, 9 Oct 2026 — github.com/QwenLM
- Hugging Face — Qwen/Qwen-Image-2.1-Turbo model card: 7B, 8 steps, 2048 default, up to 2752×1536, editing, Qwen Research License Agreement, read 11 Oct 2026 — huggingface.co
- Kandinsky Lab (GitHub) — kandinsky-6 readme: 2026/10/06 release, Lite 3B / Pro 29B, 5 s, 44 kHz, lip-sync, VSR to 1080p, performance table, 6 Oct 2026 — github.com/kandinskylab
- Hugging Face — Kandinsky-6.0-Pro-5s-Diffusers (MIT, gated) and Kandinsky-6.0-Lite-5s-Diffusers (MIT), read 11 Oct 2026 — Pro · Lite
- OpenAI — GPT-6 for everyone: GPT-6 Sol/Luna in Chat, Intelligent UI formats, 7 Oct 2026 — openai.com
- OpenAI Developer Community — GPT-6 is now rolling out to all ChatGPT users (rollout dates and tiers), 7 Oct 2026 — community.openai.com
- OpenAI — New ChatGPT ads format and measurement (visual ads during image generation; tests later in October), 5 Oct 2026 — openai.com