Atlas

Models

← All models

Gemma 3 27B IT

Google DeepMind · 2025-03-12 · 27.4B parameters

Gemma 3 27B IT is Google's open-weight, instruction-tuned multimodal Gemma 3 checkpoint, released on March 12, 2025. Its 27,432,406,640-parameter checkpoint accepts text and images and generates text with a 131,072-token context window. Post-training combines knowledge distillation from a larger instruction-tuned teacher with reinforcement learning; Google reports an August 2024 pretraining-data cutoff, while cumulative token exposure for this instruction-tuned checkpoint is undisclosed.

Benchmark scores

BenchmarkScore
A-Fantasia82.3
A-Fantasia Backwards Spelling90.0
A-Fantasia Chess72.0
A-Fantasia Cube Rotation85.0
AA-LCR5.7
AA-Omniscience Index-65.9
AidanBench2030
Aider Polyglot4.9
AIME 202520.7
Arena Text — Business, Management & Financial Ops (No Style Control)1359
Arena Text — Chinese (No Style Control)1347
Arena Text — Creative Writing (No Style Control)1345
Arena Text — English (No Style Control)1374
Arena Text — Entertainment, Sports & Media (No Style Control)1321
Arena Text — Exclude Ties (No Style Control)1322
Arena Text — Exclude Ties (Style Controlled)1332
Arena Text — Expert (No Style Control)1302
Arena Text — French (No Style Control)1369
Arena Text — French (Style Controlled)1388
Arena Text — German (No Style Control)1363
Arena Text — Hard Prompts (No Style Control)1340
Arena Text — Hard Prompts English (No Style Control)1350
Arena Text — Instruction Following (No Style Control)1322
Arena Text — Japanese (No Style Control)1287
Arena Text — Korean (No Style Control)1307
Arena Text — Legal & Government (No Style Control)1374
Arena Text — Life, Physical & Social Science (No Style Control)1383
Arena Text — Longer Query (No Style Control)1333
Arena Text — Medicine & Healthcare (No Style Control)1390
Arena Text — Multi-turn (No Style Control)1345
Loading Atlas data…