Efficiency Over Excess: Why Compressed Models Are the Key to Next-Gen NPCs
AI-generated image · Bay Street Wire
Kenji Tanaka argues that the launch of Comet 60B proves that shrinking model size, not inflating it, is the real unlock for believable, low-latency gaming agents.
For years, the industry narrative around AI has been one of sheer scale. The assumption was that more parameters equaled more intelligence, and by extension, more believable digital humans. But as someone covering the intersection of gaming and interactive tech, I believe we are hitting a wall of diminishing returns. To get to the next generation of NPCs—agents that can actually see, hear, and react in real-time without shattering the player's immersion—we don't need bigger models. We need leaner ones.
As first reported by the Financial Post, the recent announcement from Multiverse Computing is pivotal. The company has launched Comet 60B, a compressed and optimized version of NVIDIA’s Nemotron 3.5 Super VL. The technical leap here isn't about adding new capabilities, but about refining how they are delivered. The original Nemotron 3.5 Super VL utilized a multimodal Mixture-of-Experts (MoE) architecture with a total of 120 billion parameters. Comet 60B slashes that total parameter count in half to 60 billion.
In the context of game development, this is a massive win. As the Financial Post notes, the total parameter count determines the amount of GPU memory required just to load a model. For a developer trying to integrate a complex agent into a game engine that is already fighting for every megabyte of VRAM, cutting the model size by 50% is the difference between a feature that works and one that crashes the system.
Crucially, this compression doesn't come at the cost of the 'soul' of the AI. According to the Financial Post, Comet 60B maintains the same 12 billion active parameters as the original and keeps its multimodal capabilities—meaning it can still process text, images, and video. Multiverse Computing claims the model stays within 5% of the original's accuracy across these tasks, and even improves upon the original in certain benchmarks.
Enrique Lizaso, co-founder and CEO of Multiverse Computing, explained to the Financial Post that for agents to be effective, they must be able to run where the data is located while processing video across long tasks and understanding documents and images. This is the exact blueprint for the future of NPCs. Imagine a character that doesn't just recite a script, but can 'see' the player's actions via video input and respond with a level of accuracy that feels human, all while running on hardware that doesn't require a server farm in the basement.
We saw a glimpse of this trajectory in June 2026, when Multiverse Computing launched Pulsar 16B, which the Financial Post reports was built on a compressed version of NVIDIA Nemotron 3 Nano. The pattern is clear: the goal is to make advanced AI more accessible and efficient.
If we want NPCs that can handle agentic workloads—automating complex tasks and interacting across different content types—we cannot rely on bloated models. The path to believable, low-latency interactive worlds isn't through the 120-billion-parameter door; it's through the 60-billion-parameter one. By prioritizing efficiency, we finally move AI from a scripted gimmick to a living part of the game world.