Google DeepMind's Gemma 4 26B A4B — an instruction-tuned mixture-of-experts model with 25.2B total parameters but only 3.8B active per token, giving near-31B quality at MoE-efficient cost. Genuinely multimodal (text, images, and short video), with a 256K context window, native function calling, a toggleable reasoning mode, and structured output support. Scores 77.1% on LiveCodeBench v6 and 88.3% on AIME 2026. Apache 2.0. For vision requests, set max_tokens to at least 1000–1500 — the model spends part of its budget on internal reasoning before producing a visible answer.