Five open creative releases from this week, sorted by licence

None of this week's five picks is a new foundation model. Each one is machinery around a model already on the shelf, and all five are worth an hour of your time. They are listed here in the order that matters most to a working creator: whether the licence lets you sell what comes out.

That question used to be settled by whether the weights were open at all. It is not settled any more. Four of the five below are Apache 2.0 or MIT, which means the output is yours to invoice for. One is not, and it is the most downloaded item in this list. Download counts are as of September 27, 2026.

SparkWan2.1 sparse checkpoints (Alibaba)

Alibaba posted a batch of SparkDiffusion checkpoints to Hugging Face on September 24 and announced the weights on September 25. Two of them run the 14B Wan 2.1 text-to-video model in three and four passes while discarding up to 95% of attention: the 720P build is a three-step version supporting 90% to 95% sparsity, and the 480P build is a four-step version at 90%.

The checkpoints keep Wan 2.1's architecture, but they run through Alibaba's SparkDiffusion code, which wants Linux, a CUDA GPU, Triton and flash-attn, so this is not a drop-in file for an existing Wan workflow. The licence is why it is worth the setup: Apache 2.0, the same term as the base Wan 2.1 weights. The model cards list steps and sparsity but no wall-clock timings, so budget a test render before you plan a pipeline around it.

Try it

• Maker: Alibaba
• Released: posted September 24, 2026, weights announced September 25
• Base model: Wan 2.1 T2V 14B
• Licence: Apache 2.0
• Get it: https://huggingface.co/alibabagroup/SparkWan2.1-T2V-14B-720P-0.95Sparsity-3Step

ACE-Step 1.5 for Apple Silicon (roman220220)

Songs with sung, intelligible lyrics from a style description, at 48 kHz stereo, on a Mac, under MIT. The conversion went up on September 26, downloads about 5.6 GB, and generates 30 seconds of music in roughly 49 seconds on an M5 while peaking around 8.6 GB of memory.

The reason to prefer this over a hosted music tool is the licence and the training data claim: MIT on both the conversion and the base, and the ACE-Step team says the model was trained on licensed and royalty-free material. Expect to apply the two fixes in the card: the community toolkit's defaults feed the model a planner output this version does not use, and send all-zero text where the official pipeline uses a trained blank input.

Try it

• Maker: ACE-Step team, converted by roman220220
• Released: September 26, 2026
• Licence: MIT
• Hardware: Apple Silicon, about 8.6 GB peak
• Get it: https://huggingface.co/roman220220/ACE-Step1.5-sft-MLX-8bit

MicroDecoder (Vlad Mandic)

Seven files, about 68 MB, Apache 2.0, and a preview of your render in roughly a hundredth of a second per frame. It does not make generation faster. It makes a bad seed visible at step four instead of step forty, which on a slow machine is the same thing as an afternoon's worth of extra attempts.

Coverage runs across seven architecture groups including both FLUX generations, Qwen-Image-2.1, SD, SDXL and Wan 2.1, and SDNext routes to the right file automatically. The training script is published, so an uncovered architecture costs about fifteen minutes and a few hundred images rather than a wait for somebody else.

Try it

• Maker: Vlad Mandic
• Released: September 24, 2026
• Licence: Apache 2.0
• Coverage: FLUX.1, FLUX.2, Qwen-Image, SD, SDXL, Wan 2.1
• Get it: https://huggingface.co/vladmandic/MicroDecoder

Ming-Image-0.1-Design INT4 (toxicdog repack)

The 12.3 GB BF16 diffusion transformer of a 53 GB design and type model, repackaged as one 6.3 GB INT4 file for ComfyUI, MIT, uploaded on the morning of September 26. The original is InclusionAI's Ming-Image-0.1-Design, released September 17.

Clear disk space with both eyes open. The file is the diffusion transformer only: the text model, connector, MLP and VAE come from the original repository, so it is not a one-file install. The author calls it a test conversion, says output quality needs checking against your own prompts, and publishes no comparison with the BF16 original, which matters for a model whose selling point is text-rich design.

Try it

• Maker: InclusionAI, repacked by toxicdog
• Released: September 26, 2026
• Licence: MIT
• Needs: recent ComfyUI with Ming-Image and INT4 ConvRot support
• Get it: https://huggingface.co/toxicdog/Ming-Image-0.1-Design-INT4-ConvRot

Qwen-Image-2.1-viggle-turbo (Viggle)

The most downloaded item on this list, at 133,151 downloads since September 22, and the one to be careful with. Viggle distilled Alibaba's Qwen-Image-2.1 from 40 transformer passes to six with no classifier-free guidance, about 5x faster end to end, and published its own failure cases alongside the wins.

The catch is the licence. The card names the Qwen Research Licence, which permits research and evaluation only and needs a separate agreement for commercial use, and a derivative inherits its parent's terms. So this is the right tool for deciding whether the speed is real, and the wrong tool to build a client deliverable on. If you want the speed and the licence, watch which base model the compressors pick next.

Try it

• Maker: Viggle, built on Qwen-Image-2.1
• Released: September 22, 2026, current version 0.2.1
• Licence: Qwen Research Licence, evaluation only
• Claimed speed: about 5x the 40-step base model
• Get it: https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo