New AI model releases, papers, and open source: June 6-7, 2026

The June 6-7 weekend itself did not bring a net-new flagship model. What it brought was the moment a model launched a few days earlier turned into something you could actually pull down and run: MiniMax M3, an open-weight model with a 1-million-token context window and computer-use built into its weights, with downloadable checkpoints and community quantizations spreading across the early-June days. That is the release worth slowing down on, because the headline that sells it (one million tokens of context) is the exact spec your agent harness is most likely to quietly take away from you.

The headline spec is the one your harness is most likely to eat

M3 leads with a 1-million-token context window. On the model card that is a real number: the weights can attend across that much input. But a window is a ceiling, not a delivery guarantee. The amount of history the model actually receives at any given turn is decided by whatever drives the loop, and most agent runners are built to keep the token bill down by compacting older turns once the conversation gets long.

So you can download a model with a million-token window and still hand it a heavily-trimmed transcript at turn 40, because the harness decided turns 1 through 20 were old news. The model did not regress. It answered the question it was actually asked, with the context it was actually given. The window you paid for is sitting unused above the line the harness drew.

What M3 actually asks of the thing running it

Read M3's feature list as a set of requirements it places on the harness, not just bragging rights. Each property the model ships is only usable end to end if the layer above it cooperates.

The one setting that points a persistent session at your local M3

Fazm is the harness, not the model. It wraps Claude Code and Codex through the Agent Client Protocol in a native macOS UI, and it is built so the two failure modes above do not happen: sessions survive a Mac restart and auto-restore with full history, and nothing auto-compacts, so the entire conversation stays live in context for the lifetime of the window.

The concrete, checkable part is how you make that persistent session talk to an M3 you are serving yourself instead of a hosted model. There is a single field, customApiEndpoint, in SettingsPage.swift. When the value is a valid absolute http(s) URL, the bridge sets one environment variable for the agent process and disables the bundled key, so your weights and your key never leak to the wrong place:

Because the override is just ANTHROPIC_BASE_URL, anything that speaks the Anthropic request shape works behind it: a local bridge serving an M3 GGUF on your Mac, a self-hosted M3 endpoint, a corporate gateway, or a Copilot bridge. A raw OpenAI or Gemini key does not, by design, because the bridge expects the Anthropic format. And a malformed value (a bare localhost:8766 with no scheme, say) is rejected rather than silently dropped into the base URL, which would otherwise brick chat with an “Invalid URL” error on every turn.

Running M3 locally on a Mac was not a one-click weekend

Worth a caveat, because the roundups skip it. The official weights live at MiniMaxAI/MiniMax-M3 with SGLang, vLLM, and Transformers recipes, and the unsloth GGUF repo carries quantizations from 1-bit through BF16 for llama.cpp, Ollama, and LM Studio. But early in the window, llama.cpp support for M3 was preliminary and not yet in a released build, so a genuine local run meant compiling from a specific pull request, not just pulling a binary. Once you have a local endpoint up, the routing above is the same regardless of which runner you used to stand it up.

And the reach does not stop at the terminal. Whatever model you put behind the endpoint, Fazm's agent drives your actual browser and native Mac apps through accessibility APIs rather than screenshots, so an M3 you are serving locally can act across your desktop while its session stays intact across restarts and never gets compacted out from under it.

Fazm is open source at github.com/mediar-ai/fazm. Dated release windows are noisy; treat any static list as a snapshot and check the live model feeds for the current state.