Gemma 4 31B beats Qwen in Pac-Man test

The interesting part here is not the speed number alone; it is that Gemma used far fewer tokens and still delivered the more usable result. For one-shot game generation, that is often the better tradeoff because the model has to get logic, state, and interactions right in one pass.

• –Gemma finished much faster and with far less token spend, which matters on local hardware where iteration speed is part of the product
• –Pac-Man is a decent stress test for codegen because it forces pathfinding, collision handling, game state, and rendering to work together
• –Qwen’s longer output may have been more creative, but verbosity is a liability when the goal is a playable result, not just an impressive transcript
• –For local LLM workflows, “best” is often task-specific: concise correctness beats elaborate prose for gameplay code
• –The post reinforces a broader point for AI builders: benchmark models on end-to-end utility, not just raw throughput