gemma-4-26B-A4B-it

Gemma 4 26B A4B-it is Google DeepMind's instruction-tuned Mixture-of-Experts (MoE) model released in April 2026 under the Apache 2.0 license. It contains 25.2 billion total parameters with 3.8 billion active per token across 128 routed experts (8 active + 1 shared per layer), delivering near-31B dense model quality at roughly 4B-class inference compute cost. The model supports text and image inputs via a ~550M vision encoder, features a 256K-token context window, built-in chain-of-thought reasoning mode, native function calling, and multilingual support across 140+ languages.

On official Google benchmarks, the model achieves 82.6% on MMLU Pro, 88.3% on AIME 2026, 77.1% on LiveCodeBench v6, 82.3% on GPQA Diamond, and 86.3% on MMMLU. It ranks #61 on LMSYS Chatbot Arena Text with an Elo of 1438 (±8), placing it competitively among open-weight MoE models. Independent evaluations report ~76.8% on HumanEval and 68.0% on GSM8K under few-shot chain-of-thought prompting. Vision benchmarks include 73.8% on MMMU Pro and 82.4% on MATH-Vision.

The model employs a hybrid attention architecture combining 1024-token sliding-window local attention with global attention layers, unified KV cache sharing, and an autoregressive multi-token prediction drafter for speculative decoding. It is deployable via Hugging Face Transformers, vLLM, llama.cpp, and Ollama, with free API access available through Google AI Studio, OpenRouter, and Cloudflare Workers AI. Community consensus highlights it as the efficiency sweet spot in the Gemma 4 family—within ~2% of the 31B dense variant while running significantly faster on consumer hardware.