Ok, we got our first prediction wrong. Fair enough. So we're running it back: 4 predictions today, each one made by an AI running 100% on your own device with QVAC. It pulls real team stats, runs 10,000 simulations, shows its reasoning, then commits to a scoreline. Today's 4: 🇺🇸 USA v Paraguay 🇵🇾 🇧🇷 Brazil v Morocco 🇲🇦 🇩🇪 Germany v Curaçao 🇨🇼 🇳🇱 Netherlands v Japan 🇯🇵 The AI's call for each game is in the replies. How many do you think it gets right? Drop your predictions below 👇 qvac.tether.io/

Match 4 Most likely outcome: 🇳🇱 Netherlands 1 - 0 Japan 🇯🇵 Based on the 10,000 simulations: 🇳🇱 Netherlands wins 45% of the time 🇯🇵 Japan wins 29% of the time 26 % chance for a draw

How it works: A Monte Carlo model plays the match 10,000 times for the odds,using real football data. Then a local Qwen3, on-device through the QVAC SDK, reads those numbers and writes the call. It can run single matches or the whole 48-team tournament too, 5,000 times over.

We built it in 36 hours, and it is fully open source under Apache 2.0. Want to give it a try? Clone it, run it on your own machine, fork it, ship your own features. github.com/tetherto/qvac-foo…

The tournament kicks off today, so we let a local AI call the opening game. This is QVAC Football Predictor 2026: every match simulated, the local AI writes the verdict.

An image generator does not "draw." It starts from pure static, random noise with no picture in it, then removes that noise a little at a time until an image appears. That step by step cleanup is a diffusion model. Why start from noise? Because painting a full image in one shot is too hard, even for a big model. Removing a little noise is easy. Diffusion splits one impossible task into 20 to 50 easy ones, each step cleaning up the one before.

The starting static is a random "seed." Same prompt and same seed give the exact same image every time. Change the seed and you get a fresh take on the same idea. How does it learn the cleanup? Take millions of real images and add noise until each is static. Train the model to look at a noisy image and predict the noise that was added. Subtract that guess, and the picture gets one step clearer. Your prompt is a steering wheel, not a search box. At each step the model imagines the scene with your words and without them, then leans toward the version that matches. That lean is the "guidance" dial. Push it too high and images come out fried. It can run on a laptop for one reason: it never cleans up millions of pixels. It denoises a small compressed "sketch" of the image, then expands that to full resolution at the end. The whole model is a few GB, and your prompt and picture never leave your device.