Atlas

Models

← All models

Granite 4.1 30B

IBM · 2026-04-29 · 28.9B parameters

Granite 4.1 30B is IBM's 30B-class dense, decoder-only instruct language model, fine-tuned from Granite-4.1-30B-Base, with a 131,072-token context window. IBM trained the Granite 4.1 language models from scratch on approximately 15 trillion tokens, then applied supervised fine-tuning and a multistage reinforcement-learning pipeline to improve instruction following, chat, and tool calling without relying on extended reasoning traces. The weights are released under the Apache 2.0 license.

Benchmark scores

BenchmarkScore
AA-LCR18.7
AA-Omniscience Index-67.8
Artificial Analysis Agentic Index1.3
Artificial Analysis Coding Index10.4
Artificial Analysis Intelligence Index8.9
Artificial Analysis Omniscience Accuracy14.0
Artificial Analysis Omniscience Hallucination Rate95.1
Artificial Analysis Openness Index61.1
Capability125.3
CritPt0.0
GDPval-AA v2433
GDPval-AA v2 (normalized)0.0
GPQA Diamond48.1
Humanity's Last Exam (Text-Only)4.2
IFBench (Artificial Analysis)44.4
SciCode25.8
SciCode (Artificial Analysis)25.8
Terminal-Bench 2.12.6
Terminal-Bench Hard2.3
τ²-bench (Telecom)42.1
𝜏³-Banking3.9
Loading Atlas data…