Atlas

Models

← All models

Granite 4.0 350M

IBM · 2025-10-28 · 352M parameters

Granite 4.0 350M is IBM's instruction-tuned, text-only 352,379,904-parameter Nano model, fine-tuned from Granite 4.0 350M Base with supervised fine-tuning, reinforcement learning, and model merging. It is a dense decoder-only Transformer with grouped-query attention and a 32,768-token context window, offered as the traditional-Transformer alternative for runtimes where Mamba-2 support is not optimized. IBM positions the Nano family for edge and on-device deployment, and releases the weights under Apache 2.0.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-72.2
AIME 20250.0
Artificial Analysis Intelligence Index1.0
Artificial Analysis Omniscience Accuracy3.2
Artificial Analysis Omniscience Hallucination Rate77.8
Artificial Analysis Openness Index55.6
BFCL v4 Overall Accuracy19.0
Capability99.7
CritPt0.0
GPQA Diamond26.1
Humanity's Last Exam (Text-Only)5.7
IFBench (Artificial Analysis)15.9
LiveCodeBench (Artificial Analysis)2.4
SciCode (Artificial Analysis)0.9
Terminal-Bench Hard0.0
τ²-bench (Telecom)13.2
Loading Atlas data…