Why We Use the Gemma 4 Family
Strong capability at practical model sizes makes Gemma 4 a useful foundation for AeroAI's aviation model work.
For an aviation model, raw size is not the only consideration. We look for a foundation that can reason well while remaining practical to adapt and run. That balance is why Atlas-2 Core is built on Gemma 4 31B.
Gemma 4 offers strong intelligence per parameter across a range of sizes, rather than requiring one large model for every job. Google's published model card reports that Gemma 4 31B scores 85.2% on MMLU Pro and 89.2% on AIME 2026 without tools. Those are general-purpose benchmark results, not evidence that an aviation-tuned model is already accurate or ready for operational use.
The family also offers different efficiency trade-offs: E2B and E4B target on-device use, while the 26B mixture-of-experts model activates about 3.8 billion of its parameters for each inference step. The 31B dense model gives us a capable foundation for Atlas-2 Core. These choices let us consider quality, memory, and speed for different deployments without treating a single benchmark score as the whole story.
Our goal is to adapt and evaluate models against aviation-specific tasks, including reasoning, uncertainty, and responsible use of current sources. We see Gemma 4 as a strong fit for that work, not as a proven best model for every task or device. Source for model specifications and benchmark figures: https://ai.google.dev/gemma/docs/core/model_card_4
