The Small Model's Comeback

Since the release of Gemma 4, there have been accounts of how impressively capable 31B is—almost hard to believe—across general conversation, math, reasoning, and coding.

A simple experiment was done: giving the model a single image and asking it to generate a 3D model, using an image with complex geometry like an F1 car.

The comparison results were as follows.

• Claude Sonnet 4.6: The expression was flashy, but quite a few geometric anomalies appeared.
• Gemini 3.1 Pro: Cruder, but less broken.
• ChatGPT: The results were very poor.
• Qwen3.5 27B Q8: Produced its result using 6800 tokens.
• Gemma 4 31B: Produced a better output in just 3600 tokens.

The key takeaway is that Gemma 4 31B made a strong impression as a local model, both in token efficiency and output quality.