Gemma 4 26B A4B

Google · released 2026-04-01 · Apache-2.0 license

Gemma 4 26B A4B is a Google MoE open model (3.8B active of 25.2B, Apache-2.0, April 2026): 82.6% MMLU-Pro, 88.3% AIME, 77.1% LiveCodeBench, 82.3% GPQA. Trades a little quality for much faster inference than the 31B.

Key specs

TypeLocal open-weight
Parameters
Architecture
Context window
Knowledge cutoff2025-01-01
Modalitiestext, vision
Recommended backends
Minimum viable rig

Benchmark scores

GPQA Diamond82.3%
SWE-bench Verified
AIME88.3%
MMLU-Pro82.6%
BFCL v3 (tool use)
Composite score6.29
Community ratingNo reviews yet

Strengths & weaknesses

Strengths: MoE with only 3.8B active (25.2B total): near-4B speed at ~26B quality; Close to the 31B dense model at far lower inference cost; Apache-2.0, 256K context

Weaknesses: Weaker frontier reasoning (HLE 8.7) and long-context recall (MRCR 44.1); No first-party SWE-bench / tool-use scores