Gemma 4 26B A4B
Google Gemma 4 26B A4B: Mixture-of-Experts open model with 26B total but only 4B active parameters per token, running near 4B model speed with 256K context.
Model Overview
Capabilities, design details, and architectural traits
Gemma 4 26B A4B - MoE Open Model for High-Throughput Inference
Gemma 4 26B A4B is Google's Mixture-of-Experts (MoE) model in the Gemma 4 family. The A4B in its name signals its sparse activation pattern - only a subset of its total parameters activates per token during inference. Google officially documents it as running almost as fast as a much smaller dense model, while carrying the capacity of a larger model for routing and expert diversity.
| Trait | Detail |
|---|---|
| MoE inference pattern | Full set of parameters loaded into memory; only a sparse active subset used per token - explicitly designed for high-throughput with latency close to a small dense model |
| Documented speed | Runs almost as fast as a much smaller dense model, contrasting with the dense sibling which maximizes raw quality |
| MTP (Multi-Token Prediction) | Supported but MoE-specific: gains depend on batch size; at batch size 1, expert weight reuse is limited and speedups are not guaranteed across all hardware |
| Context window | 256K tokens - shared with sibling variants in the family |
| Architecture reuse | Serves as the base architecture for DiffusionGemma, Google's experimental discrete text-diffusion model |
High-Throughput Positioning within Gemma 4
Within the Gemma 4 family, the 26B A4B is the variant explicitly designated for high-throughput use. Where the dense model is positioned to maximize raw quality and fine-tuning depth, the 26B A4B trades some of that quality ceiling for significantly faster tokens-per-second in production inference workloads.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Anthropic | Anthropic | |
| Release Date | April 2, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 262K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.13 Best Input Pricing | $4 | $2 |
| Output Pricing | $0.40 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | imagetextvideo | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 13.1 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | - | - | - |
| Agentic Index | - | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Google
Other models by Google
Gemini 4 Argon
Gemini 3.8 Flash
Gemini 3.7 Flash
Gemini 3.6 Flash
Top AI Models
Leading alternatives by intelligence score
