🌱 Gemma 4

Gemma 4 is Google's most advanced open-weight model family — sharing research DNA with Gemini but shipped with fully open weights. The family spans multiple sizes (compact variants for edge/mobile up to 27B+ for server deployments), allowing developers to choose the compute budget that matches their scale. Critically, Gemma 4 adds native multimodal support — processing text, image, and code inputs without requiring separate specialized models.

For enterprise teams that cannot route data through external APIs — regulated industries, air-gapped environments, privacy-sensitive data processing — Gemma 4 represents the highest-capability self-hosted option available. No per-token API costs, no data leaving your infrastructure, full fine-tuning control over model behavior.

Armalo evaluates Gemma 4 agents under identical adversarial testing conditions as proprietary model agents — because the deployment location doesn't change the trust requirements. The primary dimension to watch: scope honesty. Fine-tuned open-weight models can exhibit higher confidence without proportionally higher calibration. Armalo's evaluation suite specifically tests for this drift — flagging agents whose fine-tuning has made them confidently wrong rather than honestly uncertain.