Gemma 4 26B A4B: Multimodal MoE for Reasoning and Coding

Google DeepMind's Gemma 4 26B A4B is a 25.2 billion total parameter Mixture-of-Experts (MoE) model, designed for efficient and powerful multimodal AI. It leverages 3.8 billion active parameters during inference, allowing it to run almost as fast as a 4B-parameter model while delivering performance comparable to larger models. This model supports a substantial 256K token context window and is proficient in over 140 languages.

Key Capabilities

• Multimodal Understanding: Processes text, images (with variable aspect ratio and resolution), and video inputs, generating text outputs. It supports interleaved multimodal input.
• Reasoning: Features configurable thinking modes for step-by-step problem-solving.
• Coding & Agentic Workflows: Achieves significant improvements in coding benchmarks and includes native function-calling support for autonomous agents.
• Native System Prompt Support: Enhances structured and controllable conversations.
• Long Context: Handles contexts up to 256K tokens.

Benchmark Highlights

• MMLU Pro: 82.6%
• AIME 2026 (no tools): 88.3%
• LiveCodeBench v6: 77.1%
• MMMU Pro (Vision): 73.8%

Good For

• Reasoning and complex problem-solving.
• Code generation, completion, and correction.
• Agentic workflows requiring structured tool use.
• Applications requiring multimodal input (text, image, video) for text generation.
• Deployments where fast inference is crucial, balancing performance with efficiency.