granite4 - 技术专题

Run frontier LLMs and VLMs locally on Qualcomm devices across NPU, GPU, and CPU with a few lines of code

A comparison of traditional RAG vs Graph RAG to see whether the results are better

mech-interp suite for Granite4 models that use Mamba-2 architecture

A terminal coding agent powered by mellea

Skirmish AI bot for Warzone2100

Systematic study of LoRA fine-tuning strategies for IBM Granite 4.0-H-Micro (Mamba-2 + Transformer hybrid). Demonstrates the impact of architecture-aware target selection and SSM core parameter co-training, including analysis of PEFT serialization behavior. Reports up to 37% relative improvement over LoRA-only baselines.