Open source AI on April 13-14, 2026, and the one env var that lets any of those weights operate a real Mac

The window covered Allen AI's OLMo 2 32B fully-open weights continuing to roll through fine-tune and evaluation forks, Mistral's Codestral 2 (Apache 2.0) picking up momentum for local coding via llama.cpp and vLLM, Meta's Llama 4 Maverick open weights driving a wave of task-specific distillations, NVIDIA's Ising family of open AI models for quantum error correction, and new inference-side updates landing in vLLM 0.8.x, llama.cpp, SGLang, and LiteLLM (the Anthropic-compatible proxy layer this page uses as its anchor). On the agent-framework side, Google ADK and Goose both pushed new open-source agent scaffolding the same week.

ANTHROPIC_BASE_URL. Fazm's Custom API Endpoint setting writes a URL into that variable on the agent bridge child process. The exact line is in Desktop/Sources/Chat/ACPBridge.swift at line 381, which reads env["ANTHROPIC_BASE_URL"] = customEndpoint. Anything that speaks the Anthropic request shape at that URL (LiteLLM with its Anthropic adapter, an Anthropic-compatible proxy in front of vLLM, llama.cpp's OpenAI-compatible server wrapped by LiteLLM's anthropic_to_openai translator, a corporate gateway) gets to handle the turn. The model on the other end can be OLMo 2 32B, Codestral 2, Llama 4 Maverick, Qwen, anything.

Three files. /Users/matthewdi/fazm/Desktop/Sources/Chat/ACPBridge.swift lines 379-382 is where the env var is injected into the ACP bridge child process. /Users/matthewdi/fazm/Desktop/Sources/MainWindow/Pages/SettingsPage.swift line 837 is where the @AppStorage("customApiEndpoint") backing property lives, and lines 904-948 render the toggle and TextField with placeholder text "https://your-proxy:8766". /Users/matthewdi/fazm/Desktop/Sources/Providers/ChatProvider.swift line 2103 is where the provider reads the same UserDefaults key when rebuilding the HTTP client on endpoint changes.

Two days before. The Custom API Endpoint setting landed in Fazm v2.2.0 on April 11, 2026. The CHANGELOG.json line reads: "Added custom API endpoint setting for proxies and corporate gateways." That is 2026-04-11, indexed by version 2.2.0. So by the time open source roundups published April 13 and 14 posts naming which weights had dropped, the consumer-side seam for routing any of those weights into a Mac agent had been shipping for 48-72 hours.

No. Fazm is a closed-source consumer Mac app. The interesting claim is narrower: Fazm composes with the open source ecosystem through an endpoint seam. The ACP bridge it starts is just a Node child process reading ANTHROPIC_BASE_URL from its environment, so a user who runs LiteLLM, a vLLM server with an Anthropic shim, or any other Anthropic-compatible proxy pointed at open weights can route a full Fazm session (chat, desktop control, memory graph, skills) through their open-model backend without Fazm ever touching the weights.

Because the roundups cover the upstream side of the stack (weights, repos, papers, benchmarks) and stop before the client. The downstream question (what does it take for a non-developer Mac user to actually use these releases for real desktop work) lives in a consumer app's Settings screen and a sixty-line LiteLLM config, not on Hugging Face or GitHub Trending. It is also invisible to anyone who has not installed the consumer app, so there is no path by which a typical roundup author would find it.

LiteLLM is the open source proxy that speaks the Anthropic request shape on one side and translates it to any provider (OpenAI, HuggingFace, vLLM, Ollama, llama.cpp server) on the other. It is the standard way to put an Anthropic-compatible endpoint in front of an open-weights model. Once you run litellm --config litellm.yaml --port 8080 locally, you set Fazm's Custom API Endpoint to http://localhost:8080, and Fazm will route every turn to whichever open model you configured in litellm.yaml. The CHANGELOG hint that this was the intended use is the phrase "proxies and corporate gateways" in the v2.2.0 entry.

Yes. Open /Users/matthewdi/fazm/Desktop/Sources/Chat/ACPBridge.swift in any editor and jump to line 379. The three-line block reads: "if let customEndpoint = defaults.string(forKey: \"customApiEndpoint\"), !customEndpoint.isEmpty { env[\"ANTHROPIC_BASE_URL\"] = customEndpoint }". Then open Desktop/Sources/MainWindow/Pages/SettingsPage.swift at line 837 to see the @AppStorage backing key, and at line 933 to see the TextField with placeholder "https://your-proxy:8766". Three files, three confirmations. Nothing on this page is invented.

Accessibility-tree text. Fazm's desktop control goes through native macOS accessibility APIs (AXUIElementCreateApplication, AXUIElementCopyAttributeValue) via a bundled mcp-server-macos-use binary, not screenshots piped into a vision model. The significance for open source release week is that the tool surface presented to the model is text, so open weights like Codestral 2 and OLMo 2 32B, which have no vision capability, can still drive Mac apps as well as any frontier model. The open-weights story is not gated on whether your model has vision.

Fazm's v2.3.2 CHANGELOG entry on 2026-04-16 reads: "Tightened privacy language in onboarding and system prompts to accurately say 'local-first' instead of 'nothing leaves your device.'" Relevant here because as soon as a user points the Custom API Endpoint at a hosted open-weights proxy (or even a localhost proxy that itself reaches out for inference), the router can leave the machine. Fully-local only holds when the ANTHROPIC_BASE_URL points at a localhost proxy backed by a localhost vLLM or llama.cpp process. Local-first is the honest default frame.