New AI model releases, papers, and open source: June 3-4, 2026

Two open-weight models landed across June 3 and June 4, 2026, and they sit at opposite ends of the size range. One is small enough to run on your laptop. The other is half a trillion parameters and was announced with a specific job in its own product copy: long-running agents. That second phrase is the one worth slowing down on, because a model built for agents that run for hours is only as long-running as the thing that holds its context.

The June 4 release names its own bottleneck

NVIDIA did not market Nemotron 3 Ultra as a chat model. The launch framing is precise: an open Mixture-of-Experts hybrid Mamba-Transformer for agentic reasoning, tool use, and long-context tasks, aimed at “long-running agents that plan, call tools, and reason across many turns.” The hybrid Mamba-Attention stack and the 550B/55B sparsity are there to make a model that stays coherent over a long run cheaper to serve.

Here is the part the spec sheet cannot fix. A model trained to reason across many turns has to actually be handed those turns at inference time. If the thing running the loop trims older messages to save tokens, or forgets the whole conversation when the app closes, the model never receives the long history it was built to use. The weights got better on June 4. Whether you feel it is decided one layer up, in the harness.

Where a long run actually breaks

Picture a single agent run of a dozen turns: read a file, call a tool, make a decision, keep going. The model is fine. The failure is structural, and it shows up at the turn where the harness decides the history is too long and quietly drops the earliest decision.

That last step is the whole problem with running a long-horizon model through a harness that compacts. The model did not regress. It was handed an incomplete transcript and answered the question it was actually asked. A restart mid-run is the same failure with a harsher edge: the session is gone, not just trimmed.

The one setting that lets the same session hold an open model

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. That is exactly the property a model “for long-running agents” needs from whatever runs it.

The concrete, checkable part is how you point that persistent session at a June 3-4 open model instead of a hosted one. There is a single field, customApiEndpoint, in SettingsPage.swift. When the value is a valid URL, the bridge sets one environment variable for the agent process:

Because the override is just ANTHROPIC_BASE_URL, anything that speaks the Anthropic API format works behind it: a local bridge serving Gemma 4 12B on your 16 GB laptop, a self-hosted Nemotron 3 Ultra endpoint, a corporate proxy, or a Copilot bridge. A raw OpenAI or Gemini key does not, by design, since the bridge expects the Anthropic request shape. The endpoint applies to Claude models, so you switch the backend without changing how you work.

Why this maps cleanly onto these two releases

The two June 3-4 models pull in opposite directions, and the harness is what makes either one usable in a real workflow rather than a benchmark screenshot.

Gemma 4 12B is the local one. It is designed to run on roughly 16 GB of unified memory and integrates with MLX and llama.cpp, which are exactly the bridges you would stand up on a Mac. The endpoint override means you can run a fully local, private agent loop: your screen and mic stay on your machine, the model stays on your machine, and the session that drives it stays intact across restarts. Reaching past the terminal still works, because the agent drives your browser and native Mac apps through accessibility APIs rather than a code-only sandbox.

Nemotron 3 Ultra is the long one. A 550B model you would self-host or hit through a hosted endpoint, built to reason across many turns. Point the persistent session at it and the run can go long without the harness eating the early decisions. If you want to compare it against your usual model on the same task, fork the chat with one click: a new window opens with the full prior context and the original is left untouched, so the two runs start from an identical history instead of a re-typed prompt.

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.