This week: Tencent's Hunyuan team open-sourced Hy-Nomas, an AI "world model" that builds a video scene a few seconds at a time and lets you change the story, or swing the camera, as it goes.

Elsewhere: How a video can remember what it already showed you, why "runs on one RTX 4090" comes with homework, and where to try it free in your browser.

The New Prime Signal

Most AI video tools work like a vending machine. You type a prompt, you wait, a clip drops out, and if the fox walks the wrong way you start over with a new coin. Hy-Nomas, a new open-source model from Tencent's Hunyuan team, works more like a choose-your-own-adventure book. You start a scene, watch a few seconds, then tell it what happens next. The model card describes it as "a video world generation model that supports free-form, multi-turn interaction through text instructions and camera control." In plain English: you get to be the director, and you don't have to yell "cut" every time you have a new idea.

What happened: Tencent published Hy-Nomas this week on Hugging Face and GitHub under the Apache 2.0 license. According to AI Lookout, both repositories were created on October 21 and the model card and weights were last updated on October 23. The release includes the model weights, inference and training code, two optional add-ons (one for a more cinematic look, one trained on real robot-arm video), and a free online demo. Tencent also released Hy-Nomas-Bench, a test suite with "280 test cases" covering text interaction, camera moves, and spatial memory. By default it outputs video at 720 × 1280 and 24 frames per second, using what the card calls "8 denoising steps" per segment, which is AI-speak for "it doesn't need to think very hard about each piece."

Why it matters: "World models" are one of the busiest corners of AI right now. They're AI systems that try to simulate a place you can move around in and change, rather than spit out a single fixed clip. They matter for games and storytelling, and for training robots in simulated worlds before they touch real ones. That's presumably why Tencent bothered training a robot-arm add-on. Many world models are locked behind company APIs. This one comes with the weights and the training code under a permissive license, so researchers and tinkerers can poke at it, retrain it, and build on it. The card also mentions a configuration that runs on "a single NVIDIA GeForce RTX 4090," a graphics card some gamers already own, though that comes with fine print we'll get to.

The fine print: First, there are no scores yet. Tencent released a benchmark but, as AI Lookout points out, published no results on it, and the GitHub page lists the technical report as "Coming Soon." So we can't tell you how well it follows instructions or how well it keeps a scene consistent. Tencent hasn't said either. Second, the single-graphics-card mode makes trade-offs. It squeezes the model into smaller number formats, keeps only a limited slice of recent history on the graphics card, and parks the rest in regular memory. The card is upfront about the cost: "System RAM must accommodate the offloaded models and the growing history archive. CPU–GPU transfers add latency." The standard setup wants eight GPUs. Third, "24 FPS" is how fast the finished video plays, not how fast it's made. The card says "Generation speed depends on hardware, resolution, segment length, history, and the selected configuration." And the online demo relies on a separate prompt-rewriting helper, Nomas-PE, that needs its own hardware.

This is a fun release, and a generous one. A world you can keep steering mid-stream is closer to how people actually make things: you rarely know the ending when you start. Giving away the weights, the training code and a test suite means others can find out how good it really is, which matters because Tencent hasn't published that yet. Treat it as an impressive sandbox, not a finished movie studio, and give the demo a spin before you decide whether it's worth a weekend on your own GPU.

The neat part isn't any single clip. It's that each new piece of video knows what came before it, without needing to know what comes next.

How it works: Hy-Nomas mixes two popular AI techniques. Inside each short segment, a diffusion model turns noise into video, the same basic idea behind most AI image generators. Across segments, it works more like a chatbot writing one word after another. Each new segment builds on everything so far. The card says "Attention is causal across segments, so future segments are not required to generate the current one." The memory trick is a cache: "A persistent key/value (KV) cache carries context between segments," so the model reuses its notes on earlier scenes instead of rereading the whole history every time. You steer it with structured prompts. The card's example starts with "A small orange fox stands on a stone path in a luminous forest. Blue fireflies drift between the trees." Five seconds later, the next instruction reads "Glowing blue flowers open beside the path as the fox continues forward." Camera paths can be fed in alongside the text, so you can pan or move through the scene too.

Try it: The easiest start is the online interactive demo, where, per the model card, you can "start a world, continue it with text instructions or camera paths, and download the generated video." If you have serious hardware, the code is on GitHub and the weights are on Hugging Face. Tencent "strongly" recommends its ready-made Docker image, because building the attention kernels by hand can take a while. The single RTX 4090 setup is there, but bring plenty of system RAM and some patience. The license is Apache 2.0, and Tencent notes that third-party pieces, such as the Qwen text encoder, keep their own licenses.