MAI Code 1.1 Flash: Microsoft brings coding AI locally — but it wants 75 GB of RAM

TVS News AI, editorially supervised, presents MAI Code 1.1 Flash: Microsoft brings coding AI locally — but it wants 75 GB of RAM.

75 gigabytes of RAM as the entry price for a coding AI that runs on your own machine — that is the announcement with which Microsoft is now making its MAI Code 1.1 Flash model available locally. Sounds at first like news for developers who don't want to send their data to the cloud anymore. And at its core, that is exactly what it is. Only the hardware hurdle is anything but ordinary.

According to Frandroid, Microsoft has introduced a local variant of its AI coding model. The model is, per Frandroid, particularly cheap and fast. Microsoft is positioning it as an alternative to cloud-based coding assistants, which typically charge monthly fees and route your inputs through outside servers.

MAI Code 1.1 Flash: What Microsoft means by "cheap and fast"

MAI Code 1.1 Flash is Microsoft's AI coding model for programming tasks — code generation, refactoring, explanations. What exactly the model delivers, you currently have to infer from what Microsoft itself communicates. Concrete benchmarks or comparison figures against other coding AIs were not to be found in the available sources.

By end of October in GitHub Copilot — but only on RTX Spark PCs

According to Frandroid, the local variant is expected to show up in GitHub Copilot by the end of October. That is the obvious distribution channel: Microsoft's own developer platform, integrated directly in the editor. But not on every PC. Frandroid explicitly writes that the model is meant to run on machines with Nvidia's RTX Spark chip — an Nvidia platform for AI workloads on the desktop, not the standard hardware in a home office.

75 GB of RAM: Why the local AI doesn't fit in every machine

The 75 GB of RAM that Frandroid cites in the headline is not a typo. Large AI models need a lot of memory because they have to keep the entire model in RAM during inference. 75 GB is, however, a lot even by AI standards. Anyone wanting to run the model locally needs either a high-end workstation or precisely that RTX Spark hardware that Microsoft is tailoring the local variant to. For casual coders, this is nothing. For development teams with sensitive projects who don't want cloud dependency, it is a potential game changer.

Claude Haiku 5.5 and GPT-6 Luna: The competition in view

Frandroid classifies MAI Code 1.1 Flash as a competitor to Claude Haiku 5.5, which, per Frandroid, has only just been released. Also named in the comparison: GPT-6 Luna. Both are cloud-based coding AIs. Microsoft's argument is apparently: same class of code assistance, but locally on your own machine. Whether the performance holds up will be shown by the first benchmarks and real-world tests.

What this means for you

If you own an RTX Spark machine or your employer has one in the server room, you'll get a locally running coding AI directly in GitHub Copilot by the end of October. If not, MAI Code 1.1 Flash is for now a story to keep an eye on — and another signal that the market for local AI models is breaking out of its niche in the coding space as well.