Three labs shipped new models in 2026. Which ones can your Mac app actually run?

The roundup, and the column the calendars leave out

Here is the 2026 release picture for the three labs people search for together. The first two columns are what every release calendar gives you. The third is the one that decides whether you can use any of it from a desktop agent today.

The release facts move week to week, so a static page is not where you track them. Public registries stay current. The durable part of this page is the third column: why a model that exists is not automatically a model your tool can run.

Why the answer splits three ways

Fazm is a native macOS app, and its agent is Claude Code, wrapped through the agent client protocol. Claude Code talks to a model in exactly one wire format: the Anthropic API. That one fact is the whole story. A model is easy to reach if it already speaks that format, and it takes a translation layer if it does not. The four model families people care about in 2026 fall on different sides of that line.

Read the diagram left to right. Whatever model you put on the left, it has to arrive at the hub in the Anthropic API format. Claude is already there. The other three need help getting there. What sits on the right, the loop and the session behavior, never changes regardless of which model fed it.

Lane 3 · the part no release calendar tells you

Gemini and Qwen: reachable, but only through a gateway

Here is the anchor fact for this whole page. Fazm has a Custom API Endpoint setting under Settings, Advanced, AI Chat. It looks like the field where you would paste a Gemini or Qwen key. It is not. When you fill it in, Fazm reads that value in ACPBridge.swift and sets one environment variable before spawning the bridge:

// Desktop/Sources/Chat/ACPBridge.swift
// Custom API endpoint (allows proxying through
// Copilot, corporate gateways, etc.)
if let customEndpoint = defaults.string(forKey: "customApiEndpoint"),
   !customEndpoint.isEmpty {
  env["ANTHROPIC_BASE_URL"] = customEndpoint
}

ANTHROPIC_BASE_URL is the override Claude Code respects for routing through a proxy. The agent still makes Anthropic-shaped API calls. It just sends them somewhere else. That is why the setting's own in-app help text is blunt about what it is not:

So a Gemini 3.1 or Qwen 3.6 release is not unreachable. It is one layer of indirection away. You run a gateway that accepts Anthropic-format requests and translates them to Gemini or Qwen calls. Open-source proxies like LiteLLM and claude-code-router do exactly this. You point the Custom API Endpoint at the gateway, not at Google or Alibaba directly, and from that moment Gemini or Qwen answers every turn while the Fazm window behaves exactly as it would on Claude.

What the field should and should not contain

How a Qwen request actually travels

Put the gateway in place and the round trip for a non-Claude model is straightforward. Every turn follows the same path. Nothing in the Fazm window knows or cares that the model on the far end is not a Claude tier.

For an open-weight Qwen model the gateway and the model can both live on your Mac, served by a local runtime like LM Studio or Ollama. Fazm has error handling tuned for that case: if the local server has no model loaded, the bridge surfaces a specific message telling you to load one rather than a generic failure. That is the whole appeal of running an open-weight release locally, and the loop never leaves your machine.

What stays constant when the model changes

The reason the protocol distinction matters less than it sounds is that the model is the cheap part to swap. The expensive part, the harness, does not move. When you point a Fazm window at Gemini through a gateway, or switch a forked window to Qwen, you keep everything that made the window useful. The session persists across a Mac restart. The one-click fork still branches a live conversation into a new window with the full prior context. The chat history stays live in context for the lifetime of the window, with no auto-compacting quietly dropping decisions you made an hour ago.

That is the practical takeaway for a year with this many releases. You do not need a new app every time a lab ships a model. You need a tool where the model is a per-window dropdown value and the protocol is handled once. If you want to compare two of these releases on a task you actually care about, the companion guide on testing a new model release on your own real work walks through the fork-and-compare flow step by step. For the open-weight side of 2026, the May 2026 open-source release notes cover what is worth serving locally.

Questions people ask about the 2026 model releases