IBM’s Granite 4.0 Nano brings open, small AI models to the edge

IBM’s AI team has released the Granite 4.0 Nano series, a set of ultra‑small language models designed for edge and on‑device workloads, expanding the Granite 4.0 family with open‑source options under the Apache 2.0 licence.

The Nano line includes ~1.5B and 350M‑parameter variants, with both hybrid and transformer architectures aimed at latency‑sensitive use cases on limited hardware.

Granite 4.0 lineup

• Granite‑4.0‑H‑1B (~1.5B parameters, hybrid SSM/transformer): positioned for edge and on‑device scenarios that need low latency.
• Granite‑4.0‑H‑350M (~350M parameters, hybrid SSM/transformer): an even smaller option to reduce compute and memory demands further.
• Granite‑4.0‑1B and Granite‑4.0‑350M (transformer baselines): alternative builds where hybrid runtimes are not yet optimised.

The Nano models sit within the broader Granite 4.0 generation, which introduced a hybrid Mamba‑2/transformer design and, in select models, a Mixture‑of‑Experts approach.

IBM stated Granite 4.0 has achieved more than 70% lower memory requirements and around 2× faster inference versus similar models, efficiencies that are pertinent for multi‑session, long‑context and edge scenarios.

Model governance and certification

IBM noted that Granite 4.0 is cryptographically signed and is, according to the company, the first open model family to carry ISO 42001 certification for AI management systems—intended to assure users about development and governance processes.

The Granite 4.0 range is presented across tiers—Small, Tiny, Micro and Nano—to match performance and cost envelopes: “Nano” is described for lightweight, local tasks where compute and connectivity are constrained; “Micro” and “Tiny” areoriented to high‑volume, low‑complexity jobs; “Small” targets stronger accuracy without frontier‑scale costs.

Training data and benchmarking claims

IBM said Granite 4.0 Nano models have been trained using the same pipelines as Granite 4.0, on more than 15 trillion tokens, and reported competitive scores versus other sub‑billion to ~1B models across knowledge, maths, code and safety benchmarks, as well as agentic tasks like instruction following and function calling.

IBM has continued to position Granite as open and enterprise‑ready. Chief executive Arvind Krishna has previously highlighted the company’s preference for open models combined with commercial tooling (watsonx), positioning openness and responsible AI as strategic priorities for customers.