Google just dropped Gemma 4... (WOAH)

Google released four model sizes — an effective 2B, effective 4B, a 26B mixture-of-experts (4B active), and a 31B dense model (1:15). This range covers everything from mobile/edge deployment to high-performance local inference on consumer GPUs, making advanced AI accessible across hardware tiers.

The 31B dense model achieves ELO scores comparable to Qwen 3.5 (397B total / 17B active) and ranks #3 on the Arena AI text leaderboard (1:52), dramatically outperforming its size class. This validates the trend that open-source models are getting smaller, better, and faster simultaneously.

The smaller 'E2B' and 'E4B' models use a novel technique called per-layer embeddings, where each decoder layer gets its own small embedding table for every token (2:50). While the embedding tables are large, they only serve as quick lookups, so the effective compute footprint during inference is much smaller than the raw parameter count.

Gemma 4 natively supports function calling, structured JSON output, and system instructions, enabling autonomous agents that interact with tools and APIs (4:25). The 31B model scored perfect on Tool Call 15 benchmarks (7:30), confirming strong real-world agent capabilities.