Prism and Kandinsky 6.0, Doom Inside Skyrim, and Nano Banana 2.1
Two new open-source video models showed up in one news cycle: Prism, a 2K-with-audio preview that wants an 80 GB card and has a disputed demo reel, and Kandinsky 6.0, which runs on a 4090. Then AI-built game mods, including my own Doom-inside-Skyrim experiment, a 0.1 Nano Banana update, and a new AI assistant called Hark.
Things are not slowing down. Two new open-source video models landed in the same news cycle, one from a pretty surprising source and doing native 2K with audio. Then AI and video games are coming together like peanut butter and chocolate, and I built a literal bash-together of Doom and Skyrim to show it. The full rundown’s in the video.
Prism: open-source video with audio, natively at 2K
Prism comes out of Tencent’s Hunyuan team, working with Fudan University per the paper — Tencent, of all places. It’s an actual new model, not another of the MiniMax H3 fine-tunes we’ve been seeing everywhere. Caveat: there’s no announcement from Tencent. The code sits under the Tencent-Hunyuan GitHub org, and the preview weights are on a personal Hugging Face account.
It generates 2K with audio natively, instead of generating around 720 and upscaling: about 10 seconds at 24 fps with 48 kHz audio, which is really nice. Oddly, image-to-video only. The demos look pretty good, though it’s probably not yet at Seedance 2.5 level, and the paper says so: the goal isn’t to be Seedance 2.5, it’s to narrow the gap.
Under the hood it’s a fine-tune of MOVA, a 32B model from January, plus an attention trick: it only generates and attends to the parts of the video where things are moving, which is how it reaches that 2K raster. I’m pretty sure we’ve heard about that one before. The catch is that this is one of the big-boy models: at 720 you’d need an 80 GB card, and that’s saying nothing about 2K. No quantized version yet.
The demo-reel dispute
Early on there was controversy: people saying the example videos were actually real videos. The authors defended themselves, saying those are generated outputs; the discussion is on the model’s Hugging Face page. I haven’t verified either way, so I’m not vouching for the clips; the ones in the video are their demo clips, shown as theirs. For now, Prism is one to watch.
Kandinsky 6.0, breaking as I recorded
Just as I was recording, a second open model appeared: Kandinsky 6.0. I haven’t had time to mess around with it, so this is a spec sheet, not a test. Unlike Prism, it’s ready to roll on consumer hardware: Lite works on an RTX 4090, and the family runs up to Pro Full HD. Per the repo’s own 4090 timings, that’s about seven minutes a clip for Lite, roughly twenty for Pro Full HD. Skipping the local install, fal already has it: image-to-video at $1.40 for five seconds.
| Prism | Kandinsky 6.0 | |
|---|---|---|
| Output | Native 2K, with audio | Up to 1080p via its upscaler, with audio and lip-sync |
| Clip length | About 10 s at 24 fps | 5 s |
| Inputs | Image-to-video only | Text-to-video and image-to-video |
| Hardware | 80 GB card for 720p; no quantized build yet | Lite on an RTX 4090 |
A tease
Something cool is coming down the pike. I can’t talk about it yet, but I got the okay to tease it, so Desktop Flamethrower Girl finally got out of the monitor. It should release sometime this week; I don’t think it’s exactly what you’re thinking, but it’s pretty cool.
Spider-Man in Gotham, Skate 3 in Call of Duty
AI and video games have started to hit their stride, moving out of AI-slop games and into real playable experiences. (For the generated-worlds-you-step-into side, see the world-model week.) Another branch is the latest models creating some really wild mods.
Spider-Man in Arkham Knight lets our boy Peter swing all around Gotham, which is delightful. Then there’s Skate 3 inside Call of Duty, which is Modern Warfare 2 running on IW4L, an LLM-written rewrite of the engine. It makes me want to scream, “Ramirez, do a kickflip.”
Doom Rim: Doom monsters in Skyrim
That inspired my own spin. I’m calling it Doom Rim, or SkyDoom — I’m actually not sure. Instead of the typical sneaky archer I always end up classed as, I am Doomguy. What’s super stupid fun is that this is legitimately Skyrim, NPCs and all, except Doom monsters spawn everywhere and I have a chaingun.
I didn’t expect this part: the Skyrim NPCs attack the Doom monsters too. Everybody is at aggro level 100% all the time, so you can’t actually talk to anyone, but they fight off the horde, and the horde can defeat them in turn. “I used to be an adventurer like you, until you took a railgun to the knee.”
The distinction that matters: this isn’t an AI-generated version of Skyrim-meets-Doom. It’s an AI-generated mod, which takes the code of both games and mashes it together, so you need to own both. I had to buy Skyrim yet again, my 13th copy. You’re welcome, Todd.
UniversalModder: the mini tutorial
UniversalModder is a skill, so you can point Claude Code, Codex, Gemini, essentially whatever you like, at it. My run: own both games, point the agent at it, say what you want. Give it a fal API key and it can generate assets (GPT Image 2, Nano Banana, Flux); I asked whether we needed that, and it said it had the general shape and didn’t. Let it grind for about an hour and you’ve pretty much got a playable game. Opus 5.5 or GPT 6.1 should handle it. A stupid game? Yes. Entertaining for about an hour? Also yes.
GTA V in a browser, Fallout New York, and the letters
The other side is games ported or spun off with AI, and I’d watch yourself here. An uncredited someone ported Grand Theft Auto V to a web browser, compressing the assets to about 700 MB. It was reportedly taken down quickly.
In a similar story, a friend of the channel, Chris First, spent about a week building Fallout New York with Claude (the press says closer to five days): a walkable Brooklyn Bridge, real quests, all in a browser. Was playable, past tense — ZeniMax, Bethesda’s parent, sent a cease-and-desist. They’re fully within their rights. If you’re thinking about a fan game, be careful when you play with other people’s IP.
Nano Banana 2.1 is out, and it’s a 0.1
The rollout was chaotic even by Google standards: it showed up in Google Flow, I ran a couple of tests, and then it got taken away. Now it’s in wide release. Google notes leaps in visual design, mask-based editing and subject consistency. But Nano Banana 2.1 is based on Gemini 3.6 Flash, so not a Pro model yet.
Community tests:
- Brant Lynch’s thriller poster. The text looks pretty good, and I appreciate the title blocks: a film by Alan Smithee, music by Tangerine Dream. Pretty nice for an 80s movie.
- Mark K’s stress test: mixed lighting, text, reflections, a crowd. A pretty good job, though the lighting is possibly a little off; I presume the table is outside, so it should carry more blue.
- A 24-person prompt. I don’t know that there are 24 people; pause and count if you like. It took the symmetry very seriously: very Wes Anderson, as is Maximus’s image, which I just like.
My own quick test: an old prompt, a photorealistic book cover for a Pacific Northwest supernatural thriller called Alan Wake, authored, as a tribute to our former generations, by Don Mindy. That’s an in-joke from, I think, 2.0, which put Mindy’s name in as garbled text. Mindy is now the official author of the Alan Wake novelization. To test text, I had it generate chapter one, and it calls out Bright Falls, the town from the game. It’s been a long time since I played it, but it feels like the opening narration.
So, look: I know a 0.1 Flash update is not the thing everyone has been clamoring for. Gemini 4 Argon is coming soon, which hypothetically means a big-brain update to Gemini image, so hopefully a big banana bread update follows. I presume it’s in the oven right now.
Hark, and every company wanting to be your AI best friend
Another AI assistant asking for all your information so it can be helpful. Hark comes from Figure founder Brett Adcock’s new company; on camera I said Figure itself was building it, and that’s wrong. It’s launching with a $100 plan, and the first 100,000 sign-ups reportedly get the paid plan free, with no duration stated. I signed up and got a free account; I haven’t run it through its paces yet. I don’t know that jumping straight to the $100 plan is the way to go.
If you haven’t played with any of these agents, you should; if the privacy side bothers you, point it at a burner Gmail account. Every company is going to be vying to be your new AI best friend. Now I’m going to have my robot butler check on the banana bread in the oven. Oh, you know what? It can’t smell, can it? Humans one, AI zero.