New AI model releases, papers, and open source: June 2-3, 2026

What shipped, at a glance

Three named models across two days, split clean down the open/closed line. The dates are the announcement dates from each vendor.

Two storylines, one two-day window

These two days are a tidy illustration of how releases actually arrive. The closed models came down a keynote-and-product channel. The open model came down a download-and-blog channel. They never sit in the same place, which is why a single dated list is always a photograph of a moving thing.

The one you can actually pull down: Gemma 4 12B

Gemma 4 12B is the release in this window that matters if you want to run something yourself. It is a single dense 12-billion-parameter decoder-only transformer that handles text, images, audio, and video natively, and the interesting design choice is that it drops the separate encoders most multimodal models carry. Per the June 3 launch coverage, vision and audio flow straight into the LLM backbone: a small 35M vision embedder does a single matrix multiplication per image patch, and raw 16 kHz audio is sliced into 40 ms frames and linearly projected into the same embedding space as text tokens. No 550M vision encoder, no 300M audio encoder bolted on the side.

Specs per the June 3 release writeup and the Gemma developer guide. Google reports it performs near the 26B mixture-of-experts model at under half the memory footprint, but did not publish the full tables on day one, so treat the headline comparison as a vendor claim until independent numbers land.

An encoder-free model reads pixels. On a Mac, that is the easy half.

Here is the part no release roundup connects. Gemma 4 12B can take an image of your screen and reason about it. That sounds like the whole game for a desktop agent, but it is only half. A raw screenshot tells a model what things look like; it does not reliably tell it which rectangle is a clickable button, what that button is labeled, or where to click so the click actually lands. Pixels are lossy about exactly the structure an agent needs to act. On macOS that structure already exists in the accessibility tree, and the harness that hands it to the model is what decides whether a multimodal release helps your day.

That distinction is exactly why the model layer changing on June 3 does not automatically change your outcomes. A better vision model with a screenshot-only harness still flails on a click target it cannot localize. The same model with an accessibility-grounded harness can read the labeled element directly. The release gives you a sharper eye; the harness decides whether the hand can reach.

How to decide if a same-week release is worth keeping

A playground tells you almost nothing about whether a model is useful for your work. A model that looks sharp on a clean prompt can fall apart the moment it carries your real files, your constraints, and the dead ends you already ruled out. The only honest test is to put it behind the same agent loop, with the same tools and the same running context, on the surface you actually work on.

Two fazm properties make that test hold over the days it takes to form a real opinion. The agent loop reaches past the terminal: through macOS accessibility APIs and the screenshot tool above, it drives your actual browser, native Mac apps, and Google Workspace, so you evaluate a model on your full surface rather than a code-only sandbox. And fazm does not auto-compact; the full chat history stays live in context for the lifetime of the window, and sessions survive a Mac restart with every window auto-restored. A trial of a June 3 open weight can span a working week without the comparison quietly drifting because a harness silently dropped earlier turns. It is fully local and open source, so nothing about the screen or mic you feed a local model leaves your machine.