Google Gemma 4 26B A4B Instruct

• 🧠 Mixture-of-Experts: ~26B total, only ~4B active per token
• ⚡ Near-dense quality with roughly 4B-parameter inference speed
• 📏 256K-token context window for medium Gemma 4 models
• 👁️ Accepts text, image, and video input
• 🔧 Native function calling and structured output
• 🧠 Configurable thinking modes for reasoning across the family
• 🌐 Multilingual support spanning many languages
• 🔒 Open weights under Apache 2.0

Google

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Gemma 4 26B A4B Instruct is the Mixture-of-Experts member of Google DeepMind's open-weight Gemma 4 family, released in April 2026. Its model card describes a sparse design with roughly 26B total parameters but only about 4B active per token, letting it run close to the speed of a small model while approaching the quality of the dense Gemma 4 31B Instruct. It targets high-throughput deployment, complementing the dense 31B variant aimed at server-grade and local use.

Against its same-family predecessor, the Gemma 3 27B Instruct, this generation broadens capabilities. Per the catalog, Gemma 4 adds configurable thinking modes for reasoning, native function calling, and structured output, and it extends the context window to 256K tokens for the medium models.

Multimodality also expands: the 26B A4B accepts text, image, and video input, and the model card notes wide multilingual coverage, all in an Apache 2.0 package. This positions it as a flexible open model for developers who want vision, reasoning, and tool use in a single deployable checkpoint.

Buyers should note that, despite the low active-parameter inference cost, all of the model's parameters must be loaded into memory for routing, so its baseline VRAM footprint resembles a dense 26B model rather than a 4B one.