DeepMind's New AI: A Gift To Humanity

TL;DR

Google DeepMind's Gemma 4 is a free, open-source AI family that runs locally on phones and even a Nintendo Switch, requiring no subscription or cloud.

Key Points

• 1.Gemma 4 runs on extremely low-end hardware. The 2-billion parameter model runs offline on a first-generation Nintendo Switch, and community members have already built offline translation, summarization, and real-time image classification apps within days of release.
• 2.The 31B dense model punches far above its weight class. It ranked #3 among open models and outperforms some models 10 times its size, using four techniques: highly curated training data, hybrid attention (local sliding window + global), native image aspect-ratio preservation, and a shared KV-cache to reduce redundant computation.
• 3.Gemma 4 excels at agentic workflows, replacing proprietary tools. It supports tool use, local coding, and can be plugged into platforms like OpenAI-compatible frontends to book flights or summarize news — users report barely noticing the difference from paid services.
• 4.The context window doubled to 256k tokens compared to Gemma 3. This handles multiple long documents comfortably, though it cannot browse the internet without an agent harness and can be confidently incorrect on real-time or highly complex tasks.
• 5.The Apache 2.0 license is the biggest upgrade over Gemma 3. Unlike Gemma 3's restrictive license that passed restrictions to derivative models, Apache 2.0 allows commercial use, modification, and derivative model creation with almost no friction — a true open-source release.