Kandinsky 6.0
6 open-weight AI video models that make sound, ranked by GPU
Kandinsky 6.0 Video leads because its 3B Lite model runs on a consumer card, while Prism needs four 80 GB GPUs for its native 2K mode.
open weights
Every SLOP TV News story about open weights.
MiniMax H3
How to run MiniMax H3 on a single 12GB graphics card
ComfyUI's repackaged weights cut MiniMax H3 from 123.6GB to 42.5GB, but the local build leaves out the hosted 2K stage behind Hailuo's showreels.
WorldPlay2
WorldPlay2 turns one image into a playable browser world
The weights landed on October 6 under a non-commercial license, and a hosted version went live on Reactor the next day.
Kandinsky 6.0
Kandinsky 6.0 Video ships open weights with 44 kHz audio
Kandinsky Lab released Lite (3B) and Pro (29B) free under MIT, with five-second clips and lip-sync that run on a 16 GB GPU.
Cloudflare
Cloudflare ships Clef, the first models it trained itself
The two open-weight models return a probability instead of a sentence, accept images and video, and are free to download under Apache 2.0.
open weights
Five open creative releases from this week, sorted by license
Five releases from September 22 to 26 built on models already out, and the license is what decides which ones can leave your machine as paid work.
Qwen-Image-2.1
Viggle distills Qwen-Image-2.1 into a six-step student
A six-step distilled student of Qwen-Image-2.1 has been downloaded 133,151 times since September 22, and the research-only license keeps the output out of client work.
ACE-Step
ACE-Step 1.5 sft gets an 8-bit Mac build under MIT
An 8-bit MLX conversion of the MIT-licensed music model sings intelligible lyrics on Apple Silicon, and its author's notes say the default toolkit settings produce cacophony.
SparkDiffusion
Alibaba's SparkWan checkpoints skip up to 97% of Wan attention
Six checkpoints on Hugging Face take the Wan 2.1 and 2.2 video models to 90%, 95% and 97% attention sparsity, all of them under Apache 2.0.
Qwen
Qwen-Image-2.1 makes transparent 2K images for research use only
The 7B model puts a real alpha channel and native 2K output in one checkpoint, while the license drops from Apache 2.0 to research-only terms.
astronex
Astronex-World 1.0 Streams Open World-Model Video on a Single 48GB GPU
The Apache-2.0 release is built on Alibaba's Wan 2.2 prior and generates 832x480 video at 24 frames per second without a render queue, its authors say.
LTX-2.5
How to Train Your Own LTX-2.5 LoRA on One GPU, Step by Step
Lightricks' open trainer turns your clips into a small adapter that teaches the 22B video model a style, a face or a camera move. Here is the whole run.
MiniMax
MiniMax shares rose 40% after it gave away H3, then fell back
The open weights logged 24.8 million downloads, but the stock now sits 77% below its March peak, and nothing proves the giveaway caused the rally.
MiniMax H3
MiniMax H3 add-ons multiply as downloads hit 4.3 million a month
September brought Video Rebirth's 8-step HyperFlow speed-up, a keyboard-driven world model called H3-World, TaoMate-H3, RefMods and a faster ComfyUI decoder.
MiniMax H3
MiniMax H3's free license excludes the US, EU, UK and South Korea
The MiniMax H3 Community License is royalty-free and allows commercial work, but it carves out four territories, a $20 million revenue ceiling and a ban on training other models.
LTX-2.5
The 5 best AI video models to run yourself: LTX-2.5 tops Wan 2.2
August's releases from Lightricks and MiniMax reshuffled the open-weights field, but license territory clauses and VRAM needs decide which models creators can actually use.