IBM's Open Source Granite 4.0 Nano AI Models Are Small Enough to Run Locally Directly in Your Browser

Why It Matters

The release delivers competitive AI performance on consumer‑grade hardware, enabling local, private inference and reducing reliance on costly cloud APIs, while expanding the open‑source, enterprise‑grade LLM ecosystem.

Summary

IBM unveiled four Granite 4.0 Nano open‑source language models—two hybrid‑SSM variants (350 M and 1.5 B parameters) and two transformer variants of similar size—available on Hugging Face under an Apache 2.0 license. The smallest models run on a modern laptop CPU with 8–16 GB RAM and can even execute in a web browser, while the 1.5 B models require only a modest GPU (6–8 GB VRAM) or sufficient system RAM for CPU‑only inference. Benchmarks show the Nano family leading its sub‑2 B class on instruction‑following, function‑calling, and safety tests, outperforming peers such as Qwen3 and Google’s Gemma. IBM positions the models for edge, privacy‑preserving, and enterprise use, and has engaged the developer community with an AMA and a roadmap for larger and specialized models.