Since 2 August the transparency obligations of the EU AI Act apply (we broke the deadlines down separately, see EU AI Act: what applies from 2 August 2026). That shifts a question many businesses have quietly ignored: not just where a model runs, but how it was built becomes auditable. Anyone who has to explain in an audit which system saw which training data faces a blank page with most open-weight models.
This is exactly where IBM's Granite family stands out. According to IBM's own announcement, Granite has been the first open model family to hold an ISO 42001 certification since November 2025. For a mid-sized company that wants to run local AI without breaking data protection rules, that is not a marketing badge. It is a piece of evidence someone else has already produced.
What ISO 42001 certifies, and what it does not
ISO 42001 is the first international standard for AI management systems. It does not state whether a single answer is correct. It states whether the provider runs a documented process for accountability, traceability, data protection and reliability. Put plainly: the workshop gets certified, not the individual product.
That sounds like a fine distinction, but it matters in practice. The EU AI Act does not demand a perfect model from deployers. It demands demonstrable diligence in how you select and operate one. A model whose origin and governance are documented and externally audited shortens that chain of evidence. You have less to prove yourself, because a recognised framework already covers part of it.
The honest caveat stays: the certificate covers IBM's management of the models, not your use of them. Your own application, your data handling and your controls remain your responsibility. What you gain is a clean starting point, not a finished compliance package.
Granite by the numbers: small enough for an SMB
Granite is deliberately not a frontier show. The current Granite 4.1 generation shipped on 29 April 2026 according to IBM, in three sizes of 3B, 8B and 30B parameters, all under the permissive Apache 2.0 license. Commercial use and local fine-tuning are allowed without asking anyone.
Two engineering choices make these models interesting for smaller firms:
- Hybrid architecture. Granite 4.0 introduced a mix of Mamba-2 layers with a few attention blocks back in October 2025, at roughly a nine to one ratio. IBM reports this cuts memory needs for long context and many concurrent sessions by more than 70 percent compared with conventional transformer models. That translates directly into cheaper hardware.
- Compact sizes. The 8B model in Granite 4.1 reaches the level of the older 32B mixture-of-experts model, per IBM, in a much simpler design. You should re-run comparisons like this against your own tasks, but the direction is clear: more capability per gigabyte.
In practice an 8B Granite in 4-bit quantisation runs on a workstation with around 8 to 12 GB of GPU memory, and the 30B model on a better equipped machine or a Mac Studio with generous unified memory. No datacentre, no eight-accelerator cluster of the kind the very largest open models now require.
Why this matters for AI Act obligations
The real value sits in the combination of local processing and documented governance. Run Granite on your own hardware and no query leaves your network. That is our core position on data sovereignty: sovereignty is about where processing happens, not whose logo is on the model. What such a setup looks like in practice is described under local AI.
The difference from any arbitrary open-weight model is verifiability. Many of the strongest open models come from labs that disclose little about training data and process (we looked at that separately, see Running Chinese open-weight models locally). For pure data processing that is secondary once you run air-gapped, as we read it. But the moment regulated output or an audit report is in play, a model with external governance evidence is easier to defend than one without.
For companies across the EU that operate high-risk or customer-facing systems, this cuts real work. A certified, open and locally run model gives your compliance officer a documented anchor instead of a shrug about where the weights came from.
Not a free pass: what you still owe
A certificate on the model does not remove your own homework. Before Granite moves into a business process, three things belong settled:
- Read the license, not the label. Apache 2.0 is uncritical and permits commercial use. Still check that the variant you use really sits under that license and not under an added condition.
- Your own tests, not vendor numbers. A small test set built around your real cases says more than any leaderboard. In your working language and your domain a model can score very differently from the English benchmark.
- A human control step where it counts. Regulated content needs a human sign-off regardless. That is a matter of process, not of the model, and the AI Act expects it.
Settle these points and you document, almost as a side effect, exactly the diligence that gets asked for when it matters.
What a first step looks like
Granite is not a plug-and-play win, but it is an unusually clean starting point for firms that take data protection seriously. The pragmatic route is a scoped pilot project: one clearly bounded use case, the smallest Granite model that solves the task reliably, and a test on your real data before you commit. If you want to know whether Granite fits your tasks and your hardware, talk to us.