Atlas

Models

← All models

Granite 4.1 3B

IBM · 2026-04-29 · 3.4B parameters

Granite 4.1 3B is IBM's 3B-class long-context instruct language model, fine-tuned from Granite-4.1-3B-Base and built as a dense decoder-only Transformer with a 131,072-token sequence length. IBM says the Granite 4.1 language-model family was pretrained on approximately 15 trillion tokens and then refined through supervised fine-tuning and multistage reinforcement learning; the 3B model is designed for instruction following, chat, tool calling, code tasks, RAG, and multilingual dialogue, without an extended-reasoning mode. Its weights are released under Apache 2.0.

Benchmark scores

BenchmarkScore
AA-LCR3.0
AA-Omniscience Index-77.4
Artificial Analysis Agentic Index0.4
Artificial Analysis Coding Index4.7
Artificial Analysis Intelligence Index4.7
Artificial Analysis Omniscience Accuracy9.3
Artificial Analysis Omniscience Hallucination Rate95.5
Artificial Analysis Openness Index61.1
Capability112.9
CritPt0.0
GDPval-AA v2127
GDPval-AA v2 (normalized)0.0
GPQA Diamond31.4
Humanity's Last Exam (Text-Only)3.4
IFBench (Artificial Analysis)33.7
SciCode (Artificial Analysis)11.9
Terminal-Bench 2.11.1
Terminal-Bench Hard2.3
τ²-bench (Telecom)19.6
𝜏³-Banking1.2
Loading Atlas data…