The Hub is quiet on fresh drops. Three older cards get a look this week. Viggle's six-step Qwen-Image student, Yandex's 80-billion-parameter MoE base, and inclusionAI's design model are the focus.
On the Hub
Viggle Qwen-Image-2.1-viggle-turbo: 6-step distilled student
Viggle posted a distilled student of Qwen-Image-2.1. The card dates from September 22. Version 0.2.1 is dated September 24. It runs text to image and instruction edits in six transformer passes instead of forty. The card claims about five times faster end to end with no classifier free guidance. LoRA rank 256 sits on the base transformer and is about 1.3 GB.
The model is a distribution matching distillation of Qwen/Qwen-Image-2.1. The Hub page shows the base relation and the diffusers pipeline. The license is qwen-research. A demo space is live for the turbo variant.1
Yandex AliceAI-Foundation-80B-A3B-Base: 80B-A3B MoE from scratch
Yandex put an Apache-2.0 MoE base on the Hub. The card was created September 12. It was last modified September 22. It is 80B total with 3B active per token and 262,144 context. The model was trained from scratch.
The README describes 48 layers with hybrid KDA plus gated attention and MoE. There are 512 experts with top-10 routed plus one shared. The card claims strong Russian factual knowledge and competitive math and code. The Hub page lists text-generation with transformers.2
inclusionAI Ming-Image-0.1-Design: 6B UI poster generator
inclusionAI released a MIT 6B text to image model for UI, infographics, posters and RGBA. The card was created September 17. It was last modified September 23. It targets text rich visual design and transparent backgrounds.
The model generates complete compositions and supports a prepended phrase for RGBA output. Inference uses the Ming-Image repo with infer.py and can be run at 2048 resolution. Prompt enhancement can use Ling-3.0-flash-VL or qwen3.8-27B. The Hub page shows the custom pipeline and MIT license.3
