Model Details
Gemma 4 26B A4B
Quick answer: Gemma 4 26B A4B is Google's October 2025 MoE multimodal model with 26B total and 4B active parameters, scoring 82.6% MMLU-Pro. It is the largest model in the Gemma 4 family, using a Mixture-of-Experts architecture for efficient inference at 4B active parameter cost.
At a Glance
Where Gemma 4 26B A4B leads
- 82.6% MMLU-Pro — strong academic knowledge with 4B active params
- MoE: 26B total, 4B active — inference cost of a 4B model, quality of a 26B
- Multimodal: text and image inputs
- Open weights under Gemma Terms of Use
- Latest Gemma generation (Q4 2025)
Where it lags
- Gemma Terms of Use — commercial restrictions
- 4B active params limits maximum throughput quality vs dense models
- Limited benchmark coverage publicly available
Best for: Google ecosystem deployments requiring strong MMLU-Pro at 4B inference cost; multimodal tasks needing efficient open-weight MoE.
What Gemma 4 26B A4B Is
Gemma 4 26B A4B is the largest model in Google's Gemma 4 family — the first Gemma generation to use Mixture-of-Experts (MoE) architecture. The "A4B" suffix indicates 4 billion active parameters per forward pass, despite the 26B total parameter count. This MoE approach delivers higher model capacity at lower inference cost.
The 82.6% MMLU-Pro score is notable for a model with 4B active parameters — comparable to models that use 13-27B dense parameters. This demonstrates the efficiency benefits of MoE training: broader knowledge with manageable inference compute.
Gemma 4 26B A4B targets teams who need stronger capabilities than Gemma 3 27B (67.5% MMLU-Pro) while maintaining on-device or efficient inference characteristics. The multimodal capability extends to image understanding alongside improved text performance.
Specifications
| Field | Value |
|---|---|
| Organization | |
| Total parameters | 26B (MoE) |
| Active parameters | 4B per forward pass |
| Context window | 128,000 tokens |
| License | Gemma Terms of Use |
| HuggingFace | google/gemma-4-26b-a4b-it |
| Release date | October 2025 |
| Knowledge cutoff | June 2025 |
| Modality | Text + Vision (multimodal) |
| Architecture | Mixture-of-Experts (MoE) |
Pricing
Open weights under Gemma Terms of Use — free to self-host.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| MMLU-Pro | 82.6% | Benchgen evaluation | 2025-10 |
Gemma 4 26B A4B vs Alternatives
| Model | MMLU-Pro | Active Params | Vision | License |
|---|---|---|---|---|
| Gemma 4 26B A4B | 82.6% | 4B | Yes | Gemma ToU |
| Gemma 3 27B | 67.5% | 27B (dense) | Yes | Gemma ToU |
| Gemma 4 E4B | 69.4% | ~4B | Yes | Gemma ToU |
| Llama 4 Maverick | 80.5% | 17B active | Yes | Llama 4 |
Gemma 4 26B A4B vs Gemma 3 27B: +15.1pp MMLU-Pro (82.6% vs 67.5%) at same inference cost tier — a significant generation-on-generation improvement.
Related Models & Benchmarks
- Compare: Gemma 3 27B · Gemma 4 E4B · Llama 4 Maverick
- Benchmarks: MMLU-Pro
Frequently Asked Questions
What is Gemma 4 26B A4B?
Google's October 2025 MoE multimodal model with 26B total and 4B active parameters, scoring 82.6% MMLU-Pro. The latest and most capable Gemma 4 model with MoE architecture.What does A4B mean in Gemma 4?
A4B = 4B Active parameters. In MoE models, only a subset of total parameters are used per forward pass. Gemma 4 26B A4B uses 4B active parameters (matching inference cost to a 4B dense model) while having 26B total trained weights.Should I use Gemma 4 26B A4B or Gemma 3 27B?
Gemma 4 26B A4B for new deployments: +15.1pp MMLU-Pro, MoE architecture for efficiency, more recent knowledge cutoff (June 2025 vs September 2024).Specs from Google's official Gemma 4 release (October 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.