Microsoft is demonstrating its new WHAMM model with a real-time generative version of Quake II. You can play the AI-generated game in your browser, although it’s really just a proof of concept—it’s not very fun.
Real-time generative gaming will destroy or uplift the gaming industry, depending on who you ask. In any case, development of generative gaming models is accelerating at a rapid pace. Researchers have figured out how to run Doom within Google’s GameNGen neural learning model, OASIS AI lets you play an insane generative version of Minecraft in the browser, and in February of 2025, Microsoft introduced its unique WHAM generative gaming system.
The WHAM-1.6B model that Microsoft showed off six weeks ago was impressive but impractical. It generated a single frame per second at 300 x 180 resolution, and it required seven years worth of training data to create a “playable” game. Microsoft began work on the upgraded WHAMM (World Human Action MaskGIT Model) shortly after debuting WHAM-1.6B, and the results are kind of shocking. Not only does WHAMM work at a 600 x 340 video resolution (twice that of WHAM-1.6B), but it outputs images at a minimum 10 FPS and managed to copy Quake II with just one week of training data.
A refined architecture is responsible for these improvements. Instead of using a typical autoregressive “LLM-like” system where the AI model generates one token at a time, WHAMM’s MaskGIT setup can “generate all of the tokens for an image in as many generations as we want.” In other words, the new model utilizes parallel processing to boost output speed, image quality, and predictive accuracy.
Games generated by WHAMM are, from a practical standpoint, not very fun to play. It looks blurry, smudgy, and crusty, the frame rate isn’t ideal (though it’s not all that different from what gamers experienced in 1997), and in-game enemies are practically unrecognizable. The demo is excruciatingly laggy, too, though Microsoft blames the “noticeable latency” on its web player, rather than the model itself.
WHAMM also suffers from the “short-term memory” problem that we see in other generative gaming models. As a predictive model, WHAMM generates new frames by looking at previous frames—it’s bad at keeping track of health and ammo, enemies may disappear if you look away from them (or randomly appear for no reason), and if you push your character against a wall or stare at the floor, you may be teleported to a different location on the map.
However, in my testing, WHAMM seems to have fewer “short-term memory” problems than some other models. Its 0.9-second context length is just good enough to prevent the brain-bending, trippy craziness that I experienced when playing with generative Minecraft, although context length is clearly a huge challenge that Microsoft will need to overcome.
I should also note that WHAMM was only trained on the first level of Quake II. If you get on the elevator at the end of the level, the model freezes. So, Microsoft’s assertion that WHAMM can be trained on a weeks’ worth of video data is kind of misleading—the model requires less training data than WHAM-1.6B, but the amount of data required to generate an interactive game will vary based on content length, game complexity, and other factors.
As for how this technology will be used in the future—well, Microsoft knows that real-time generative AI can produce “new kinds of interactive media,” but it’s still exploring what that media should be.
You can play the real-time generated version of Quake II at the Copilot Labs website. Games are timed and will reset when the timer runs out. Again, this game is just a proof of concept, so don’t expect too much from it.
Source: Microsoft
