Model

Gemma 4 26B A4B

OPEN

Google · Gemma 4 family · released Apr 2, 2026

The MoE sibling of the dense Gemma 4 31B, flagged as missing in the dense model's own verificationNotes and now confirmed directly: 128 experts, 8 active + 1 shared per token, near-31B quality while running about as fast as a 4B model.

ReasoningCodingVisionFunction callingTool useAgentic

1819.4

Elo · rank #79

Parameters

25.2B

Active params

3.8B (MoE)

Context

256K tokens

Architecture

Sparse Mixture-of-Experts Transformer — 128 total experts (8 active + 1 shared routed per token), 30 layers, 1024-token sliding-window local attention, multimodal text+image encoder (~550M params)

License

Apache-2.0

Languages

140+

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath88.3%#50

best: GPT-5.2 · 100.0%

CodeforcesCoding1718#31

best: DeepSeek-V4-Pro (Think Max) · 3206

GPQA DiamondReasoning82.3%#68

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning8.7%#93

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning72.0%#22

best: MiniMax M3 · 83.0%

IFEvalReasoning98.5%#1

best: this model · 98.5%

LiveCodeBenchCoding77.1%#55

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision82.4%#9

best: Seed 2.1 Pro · 92.6%

MMLU-ProKnowledge82.6%#44

best: Claude Fable 5 · 91.5%

MMMLUKnowledge86.3%#16

best: Gemini 3.1 Pro · 92.6%

MMMU-ProVision73.8%#28

best: Claude Opus 4.7 · 85.5%

τ²-Bench AirlineAgents76.0%#2

best: Trinity-Large-Thinking · 88.0%

τ²-Bench RetailAgents85.5%#5

best: Claude Sonnet 4.6 · 91.7%

τ²-Bench TelecomAgents43.0%#33

best: Claude Opus 4.6 · 99.3%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Gemma 4 family

Elo progression across releases

Gemma 4Apr 20261997.8 Gemma 4 26B A4BApr 20261819.4

Resources

HFgoogle/gemma-4-26B-A4B-it DOCShttps://ai.google.dev/gemma/docs/core/model_card_4 APIAvailable via Google API

API price weights · each benchmark row carries its own source badge (see methodology)