Today's tools make it easy to construct AI-powered applications. But one complex area that the majority, if not all, developers wish to avoid is tips on how to host the models they use. It's one thing to make a choice from OpenAI's GPT-4o, Meta's Lllama 3, Google's Gemini, or the various open source models available on the market. It's quite one other to deploy it.
Such obligatory but tricky work could frustrate developers and distract them from their entrepreneurial ideas. But Microsoft has an answer that might make it easier to focus more on the creative process than on maintaining the models. It's called Models-as-a-Service (MaaS), and it's the AI equivalent of cloud services, charging for access fairly than infrastructure. It's available through the corporate's AI Azure Studio product.
Keep it easy
“If you've ever tried to deploy a model, there are numerous mixtures of incantations and Pytorch versions and CPU and GPU stuff,” Seth Juarez, principal program manager for Microsoft's AI platform, told VentureBeat. “Models-as-a-Service form of abstracts all of that in order that if you need to use a model and it's open source or built by OpenAI, we provide you with a catalog. You press a button and now you’ve gotten an endpoint to make use of it.”
Developers can rent inference APIs and host fine-tuning through a pay-as-you-go plan—all without having to make use of a virtual machine. Juarez explains that while Microsoft has 1,600 models for various tasks, it also desires to make it easier for developers to integrate AI functionality into their software, and MaaS is one method to achieve that.
From its launch in 2023 to today, Microsoft has made select models available through this program. First available were Mistral-7B and Metas Llama 2. This week, Nixtla's TimeGen-1 and Core42 JAIS were added, and it says those from AI21, Bria AI, Gretel Labs, NTT Data, Stability AI, and Cohere shall be available soon. This is only a small portion of what is offered on AI Azure Studio. So how do you grow to be a MaaS model?
Some are the results of corporate partnerships, and Juarez admits he doesn't understand how that happens. Others are supported because some API work has been done to make the function signatures of those models uniform enough to be a part of Models-as-a-Service. There's a uniform method to access those models. Unfortunately, more specialized models aren't allowed and have to be deployed in other ways. “That's why some are enabled as Models-as-a-Service and others may be moved into your individual container and run in what we call managed inference,” he says.
How to “rent” or “own” your models
He believes that in the longer term we’ll see a forked paradigm where developers will select models in a way that is analogous to a home-owner or renter. “Basically, you own the container, the model and Azure ML, and also you pay the rent and form of care for the upkeep,” Juarez notes. “With Models-as-a-Service, we care for the upkeep. And the more of those models we put on the market, the higher if you need to rent them. But there are also other people who find themselves very specific behind a virtual network and wish to run things on it.”
MaaS will not be a novel … model. But what happened that AI became the important technology mimicking the cloud computing business? Juarez says the established order is reversed — tech corporations aren’t any longer pushing for the technology they think we’d like. Now we're demanding features and services from tech corporations. This is because AI research and commercialization are moving almost in lockstep. “At least for my part, that's what's causing this weird inversion where consumers are demanding this type of experience by utilizing ChatGPT. And now the corporate is attempting to catch up … the user is demanding the research experiences today.”
