Microsoft has published Mage, a two-model research family that keeps every component under a fixed 4 billion parameters. The project first surfaced at AI Insiders as a compact image generator; the version now public splits that work in two, adding a dedicated video and image understanding model alongside the generator. The split is the news: one model reads, one model draws, and both stay small enough to run without a data center.

Mage-VL handles understanding. Instead of treating a video as a flat stack of still frames, it processes footage the way a video codec does, splitting each clip into anchor frames and predicted frames built from 16x16 patches. Microsoft describes the design as bio-inspired and calls it proactive streaming, meaning the model consumes video continuously rather than waiting for a clip to finish before reasoning about it.

That codec framing matters beyond terminology. A codec only stores full detail for anchor frames and encodes the frames between them as changes relative to that anchor, because most of a video’s information does not shift much frame to frame. Applying the same logic to a vision-language model lets it skip redundant computation on visual content that already repeats, freeing capacity for the visual signal that actually changes. For a model capped at 4B parameters, that is not a stylistic choice. It is how the family fits video reasoning into a budget that would otherwise only support still images.

Mage-Flow, the generation half, turns text prompts into images and reworks existing ones from written instructions, rendering output at whatever resolution the source image actually uses rather than forcing a fixed square crop. It pairs a component Microsoft calls Mage-VAE with a custom diffusion transformer, and it ships in three variants: a Base model, a reinforcement-learning-tuned version, and a 4-step Turbo build for faster inference.

Microsoft’s project page claims the family holds its own against far bigger open-weight rivals, but it publishes no benchmark scores, leaderboard placements, or comparison charts to support that claim. Treat the competitiveness claim as a company assertion until independent evaluations appear.

The fixed 4B budget is the operator signal. Microsoft is explicitly building for teams that cannot assume access to a large GPU cluster: researchers, startups, and independent developers who need a model that trains and fine-tunes on a single workstation-class setup rather than a multi-node training run. That audience, not enterprise buyers chasing frontier scores, is who a project like this actually serves.

Teams evaluating small vision-language models for video understanding should pilot Mage-VL against their own footage before committing, since Microsoft has not published third-party benchmark results to validate the codec-native approach at scale.

Microsoft’s Mage Team published the project details on its official Mage project page (microsoft.github.io/Mage).