Atlas

Models

← All models

Gemma 4 26B A4B

Google DeepMind · 2026-04-02 · 26.5B parameters

Gemma 4 26B A4B is the pretrained base checkpoint of Google DeepMind's multimodal Gemma 4 mixture-of-experts model, released alongside the family in April 2026. The checkpoint accepts text, images, and video frames over a 262,144-token context window and generates text; Google reports approximately 3.8 billion active parameters during inference.

Benchmark scores

BenchmarkScore
ALE-Bench927
Arena Code WebDev — Brand & Marketing1396
Arena Code WebDev — Consumer Product1343
Arena Code WebDev — Content Creation Tools1324
Arena Code WebDev — Data & Analytics1344
Arena Code WebDev — Gaming1359
Arena Code WebDev — HTML1375
Arena Code WebDev — React1359
Arena Code WebDev — Reference-Based Design1368
Arena Code WebDev — Simulations1332
Arena Text — Business, Management & Financial Ops (No Style Control)1435
Arena Text — Chinese (No Style Control)1478
Arena Text — Creative Writing (No Style Control)1406
Arena Text — English (No Style Control)1443
Arena Text — Entertainment, Sports & Media (No Style Control)1407
Arena Text — Exclude Ties (No Style Control)1430
Arena Text — Exclude Ties (Style Controlled)1436
Arena Text — Expert (No Style Control)1448
Arena Text — French (No Style Control)1467
Arena Text — French (Style Controlled)1478
Arena Text — Hard Prompts (No Style Control)1440
Arena Text — Hard Prompts English (No Style Control)1446
Arena Text — Instruction Following (No Style Control)1423
Arena Text — Legal & Government (No Style Control)1457
Arena Text — Life, Physical & Social Science (No Style Control)1443
Arena Text — Longer Query (No Style Control)1430
Arena Text — Medicine & Healthcare (No Style Control)1424
Arena Text — Multi-turn (No Style Control)1440
Arena Text — Non-English (No Style Control)1424
Arena Text — Non-English (Style Controlled)1425
Loading Atlas data…