Atlas

Models

← All models

Granite 4.0 H 350M

IBM · 2025-10-28 · 340M parameters

Granite 4.0 H 350M is IBM's 340,332,224-parameter dense hybrid Mamba-2/Transformer instruction model in the Granite 4.0 Nano family, with a 32,768-token context window. It was fine-tuned from Granite 4.0 H 350M Base using supervised fine-tuning, reinforcement learning, and model merging, and is intended for edge and on-device use cases including instruction following, tool calling, RAG, coding, and multilingual dialogue. The weights are released under Apache 2.0; IBM states that the AI management system for its Granite language models is certified to ISO/IEC 42001:2023.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-87.3
AIME 20251.3
Artificial Analysis Intelligence Index1.0
Artificial Analysis Omniscience Accuracy3.7
Artificial Analysis Omniscience Hallucination Rate94.4
Artificial Analysis Openness Index55.6
Capability97.3
CritPt0.0
GPQA Diamond25.7
Humanity's Last Exam (Text-Only)6.4
IFBench (Artificial Analysis)17.6
LiveCodeBench (Artificial Analysis)1.9
SciCode (Artificial Analysis)1.7
Terminal-Bench Hard0.0
τ²-bench (Telecom)14.6
Loading Atlas data…