Your next AI renewal now has a new line in the negotiation. Microsoft has said trusted customers will be able to get access to its own model weights, and a weights option changes what you can ask every other vendor to prove on price, data terms and control.
Direct answer: Microsoft announced MAI Licensed Weights, under which trusted customers will be able to get access to Microsoft AI model weights through its Frontier Co. For a Canadian firm, the first effect is a contract effect. Ask your current vendor what happens to your price and your data terms if a licensed weights option sits on the table, check where a weights based model would run and who holds the keys, and price the staff time to run a model yourself before you treat it as a saving.
Microsoft describes the MAI family as compute efficient in pre-training and token efficient in product deployment, with a goal of putting additional points on the frontier curve informed by real world task performance. Test those claims on your own work. A customer who can hold the weights negotiates differently from one who can only rent access.
What else did Microsoft put on the table?
Code 2 is the flagship family for agentic coding, presented as strong on taking code to production, deep debugging, multifile implementation, repo wide reasoning and user interface work at a fraction of frontier cost. Treat that cost line as a claim to test on your own repository. Code 2 Flash covers fast high volume work such as code explanations, targeted edits and routine fixes at much lower cost. Both are generally available on GitHub Copilot, through the model picker and Copilot Fusion, and in Foundry.
Cyber 2 Flash is described as Microsoft’s most capable security model yet, a flash sized model ahead of frontier models at a fraction of the cost per task. That is also the maker’s claim. It was co-designed with the tools where it runs and is integrated with Perception.
Microsoft-Decision-1.1 is generally available with multimodal support. More than 4,000 customers use it, and Microsoft uses it itself for routing, data labelling, personalisation and content filtering among hundreds of use cases. GPT 6.2 and Meta’s latest models are both announced as coming to Foundry. The keynote remarks also set out Agent 365 Business Analytics as a way to see what drives AI spend and connect spend to outcomes, an idea described as an ERP system for AI usage.
Why does a weights licence change a Canadian contract talk?
Most Canadian firms buy AI as a service. The vendor holds the model, sets the meter and writes the data terms. A licensed weights option adds a second shape, where your firm, or a provider for your firm, runs a model whose weights you hold under licence.
That shape raises 3 money questions. Price comes first. If a weights option exists, a service only vendor has to show what its margin buys, in support, updates and liability, rather than in access alone. Data comes second. A model running where your firm controls gives you a clearer line on where records rest and who can read them. Running cost comes third. Weights remove one bill and add others, in compute and staff time.
The test in Build or buy your next AI tool. The premium has to earn its place applies here. A premium earns its place where the vendor carries work your firm cannot staff. It loses that place where the vendor sells access and little else.
What should you ask before the next renewal?
Send the questions in writing and file the answers beside the quote. The table is the kept asset from this piece. Use it with your current vendor first.
| Question | What a clear answer shows |
|---|---|
| What happens to our price if a licensed weights option sits on the table? | Whether access alone still carries the current margin |
| What happens to our data terms under each option? | Retention, deletion and use of your records in plain writing |
| Where would a weights based model run? | The hosting place, the region and the account that holds it |
| Who holds the keys and the access logs? | Named control your firm can audit |
| What staff time does running the model take each month? | The internal cost in hours, counted before any saving is claimed |
| What does the exit look like at the end of the term? | Your data and settings back in a readable form |
The audit habit in One prompt finds every AI tool you are wasting money on gives you the starting inventory. Run the table on your largest lines first, while the renewal still gives your questions weight.
Where does the Canadian angle bite?
Canadian firms work under Canadian privacy duties and contracts priced in CAD. A hosting answer that names a place and a key holder is worth more here than a general promise about security. Ask which region holds the data and what proof you get when you ask who accessed the system.
Retail shows how fast these choices reach the customer. Loblaw is the first Canadian retailer inside Muse. The basket is the new storefront put a Canadian brand inside a model run by someone else. A weights option gives a firm of that scale a second path, with more control on one side and more running work on the other. Smaller firms face the same trade, and the staff line decides it more often than the licence line.
Price the people before you price the model. A weights based setup needs a named person who patches it, watches its costs and owns its failures. If that person is not on your team, compare a managed provider against your current service, with both quotes in CAD and both staff loads written down.
Close: The checked figure is a working one. More than 4,000 customers already use Microsoft-Decision-1.1, across hundreds of use cases at Microsoft itself. So act on the dated step. Send the 6 vendor questions in the table to your current vendor before your next renewal date, file the answers beside the quote in CAD, and let the price, data terms, hosting place, key holder and staff hours decide whether licensed weights earn a place on your vendor list.