Tencent Hunyuan and Fudan put out a video and audio model that generates both in one pass, and the headline all day has been native 2K under an MIT licence. We read the licence file and listed the repository. The licence is better than people are saying. The checkpoints are smaller than people are saying. Nobody has run it yet.
I ran a tiny print shop on one RTX 5090. Ming Image 0.1 made the sign, the price board, the stickers and the poster, Qwen Image 2.1 printed our real logo onto my shirt, and MiniMax H3 had me talk in it. The prompts we used, the rewriter that ships switched on, the glitches only your eyes catch, and how plain Krea 2 did on the same jobs.
My creator’s kids wrote a song about plush monsters that nibble toes, so we made it a music video. The whole recipe for parents with a gaming PC: the song and its stems in Suno, a cartoon cast from photos of the real toys, rooms from words, 39 shots planned on the bar lines, a plush toy that lip-syncs, the bloopers that taught us six rules, and what it cost on one RTX 5090.
Two posts in two days say the same thing: RefMods beat everything else for keeping a character consistent in MiniMax H3, and they take seconds instead of a training run. Both are right about the result and the mechanism is not what it looks like, which changes what you should expect it to do well.
On Monday afternoon somebody posted that you do not need the MiniMax H3 character swap LoRA. Thirteen hours later somebody posted that you do. Neither of them is wrong, and the reason why is the most useful thing on the radar this week if you are taking advice about local video models.
The same sentence, the same photos, four free AI photo editors on one RTX 5090. Which kept the face, which smoothed away the freckles, which re-posed the person, the prompt enhancer that leaks its own reasoning, the one-view prompt that gets a side profile in 12 seconds, and the resolution setting that makes edits nine times slower.
Someone opened a tab on Monday to try a ComfyUI node pack, came back on Tuesday, and the repository was a 404. Minutes later the author’s entire GitHub account was gone. The model weights it wrapped are still sitting on Hugging Face, untouched. Most people back up the wrong half of their install, and we were doing it too.
A 31-second photoreal film of a consistent AI presenter, made on one RTX 5090 with no cuts: reference pictures given jobs, a locked voice track, three cheap seed previews then one upscaled finish, each clip extending the last. The five ways it broke, the picture that fixed each one, the plate rules, the LoRA retrain, and the bill.
Qwen released a new image model today at 7.1 billion parameters, roughly a third the size of everything else in the line, with ComfyUI support the same afternoon. Every earlier Qwen image release was Apache 2.0. This one is research and evaluation only, and we read the licence file rather than the thread about it.
Somebody asked today how you count the real cost of AI video once the failed generations are included. We published a 16 minute episode this morning, we keep every take we render, so here are our actual numbers: what we made, what we used, what the retries cost, and the line item nobody prices.