Local Minimax M2.7, GTA benchmark
The new 2.7B parameter model creates interactive 3D worlds with cars, trees, and flocking birds from simple text instructions.
Minimax's latest small language model, M2.7, is making waves by demonstrating an unexpected aptitude for generating complex, interactive 3D applications from simple text prompts. In a viral test, a user asked the 2.7 billion parameter model to "create a 3D GTA-like experience in a single web page" where a player could walk around and drive cars. The model successfully produced a functional prototype with basic mechanics. When given iterative feedback—such as correcting reversed car light placements and control directions, and adding environmental details like trees and flocks of birds using the boids algorithm—the model coherently implemented the changes. This was achieved while the model was running in a highly efficient, quantized state (IQ2_XXS) for maximum speed, highlighting its performance efficiency.
This test, conducted outside of a formal agentic framework in a simple OpenWebUI artifacts window, underscores the model's robust coding and spatial reasoning capabilities. While competitors like GLM 5 might still lead in aesthetic polish and adding unprompted detail, Minimax M2.7 shows impressive competence in understanding and executing complex, multi-step instructions for game development. The user also noted its strong performance in other coding tasks within the OpenCode environment. The demonstration suggests that smaller, efficiently quantized models are becoming increasingly capable of handling creative technical tasks that were previously the domain of much larger AI systems, potentially lowering the barrier to rapid prototyping and interactive content creation.
- Minimax's 2.7B parameter M2.7 model generated a playable 3D GTA-style game from text prompts in a single HTML file.
- The model iteratively improved the game based on user feedback, adding trees and implementing a boids algorithm for flocking birds.
- It performed this complex task while running in a highly efficient IQ2_XXS quantized state, prioritizing speed without losing coherence.
Why It Matters
It demonstrates how small, efficient AI models can rapidly prototype complex interactive applications, lowering the barrier for game and simulation development.