IBM's Granite 4.2 family is one of the more generous open-weight releases of 2026. It ships in three sizes, 3B, 8B and 30B parameters, all with reasoning built in, and all under the Apache 2.0 licence. There is no usage fee, no commercial-use restriction, and the smaller weights will run on a single decent GPU or even a capable laptop.
That generosity leads a lot of Australian business owners to a quick conclusion: the licence costs nothing, so the model costs nothing. The licence is real. The conclusion is not. This piece walks through where the money actually goes when a mid-market business puts Granite 4.2 into production, and the narrow case where it still makes sense.
Is IBM Granite 4.2 actually free for commercial use in Australia?
Yes, the licence is free. Apache 2.0 allows commercial use, modification and redistribution with no fee and no user cap, so an Australian business can download Granite 4.2 and build it into a product or internal tool without paying IBM anything. What the licence does not cover is the hardware, people and process needed to run the model reliably, and those costs are where the real bill sits.
That distinction matters because the licence line is the one number on the model card that looks final. Everything else (hosting, monitoring, tuning, upgrades) is left for you to estimate, and most early estimates leave it out entirely.
Where the money goes after the download
For a 30B model running a real workload, the costs that decide your total sit downstream of the download:
GPU hosting. A 30B model comfortably needs a mid-tier GPU instance running continuously, typically $500 to $1,500 a month depending on region and redundancy.
Operations and monitoring. Someone has to watch for drift, patch the serving stack and handle outages. For most SMBs that means a contractor or a slice of internal engineering time.
Fine-tuning and evaluation. Getting a general reasoning model to behave well on your own documents or tickets is rarely a weekend job.
The next upgrade. Granite point releases arrive every few months, and each one is another qualification cycle before it touches production.
We have costed the people side of this in detail in our DevOps labour breakdown for Australian teams, and the hardware side in the 2026 self-hosting cost numbers. The short version is that labour, not GPUs, tends to be the larger and less predictable line.
| Cost line | What drives it | Illustrative annual (AUD) |
|---|---|---|
| Licence | Apache 2.0 | $0 |
| GPU hosting | Continuous mid-tier instance | $6,000 to $18,000 |
| Ops, monitoring, patching | Contractor or internal time | $12,000 to $25,000 |
| Tuning and evaluation | Initial fit to your data | $5,000 to $12,000 |
| Upgrade qualification | Each point release | $5,000 to $10,000 |
| All-in range | Typical outcome | $28,000 to $65,000 |
The rows are illustrative and will not add neatly for every firm, but the all-in band of $28,000 to $65,000 a year is what we usually see for a 15 to 40 person Australian services business once contractor time, GPU rental, monitoring and the occasional outage are counted. That is the number to put beside the $0 licence.
How the managed alternative compares
The comparison that matters is not Granite against nothing. It is Granite against a managed model doing the same job. For most mid-market firms that comparison has three parts:
Usage cost. A Claude API bill scaled to the same work typically comes in lower for businesses using under a few million tokens a day, and it carries no infrastructure headcount.
Upgrades. Managed model updates arrive without a separate qualification project every time a new version ships.
Compliance. A managed platform with enterprise controls is usually a faster answer in an APRA or Privacy Act review than a self-hosted model and its patch log.
If your volume is genuinely high, the break-even shifts. Our token-volume break-even maths shows how to find that crossover for your own usage, and the ROI calculator will give you a first-pass figure in a few minutes.
When Granite 4.2 belongs on the shortlist
None of this makes Granite 4.2 a poor model. It earns a place in one specific scenario: a workload that genuinely cannot leave your premises or your own cloud tenancy, where the reasoning task is narrow enough that a 30B model handles it well. A records-classification job inside a sovereign environment is a reasonable example. A general-purpose staff assistant is usually not.
The 3B and 8B sizes widen that case a little. A small model running on existing hardware for a single, well-defined task (tagging, routing, extraction) can avoid the dedicated GPU line altogether. The ops, evaluation and upgrade costs still apply, though, so the saving is smaller than the hardware difference suggests.
A quick test before you commit
Ask three questions. Does the data have a hard requirement to stay on infrastructure you control? Is the task narrow and stable enough to evaluate properly? Do you have, or will you pay for, someone who owns the model in production? If the answer to any of these is no, the managed route is usually cheaper in practice.
What not to conclude
This is not an argument that open weights are always the expensive option. Firms with existing GPU capacity, an in-house ML team and a sovereignty requirement can run Granite 4.2 well and at a fair price. Nor does a managed model remove cost discipline: API bills grow with usage and need the same governance as any other cloud spend. The point is narrower. A free licence answers the licensing question and nothing else, so run the same total-cost-of-ownership exercise you would for any build-versus-buy decision.
If you want help running that comparison against your own workloads, get in touch and we can work through the numbers with you.



