Local LLM news, April 2026: one environment variable is the whole thing

Alibaba's Qwen 3 family (0.6B to 72B, dual thinking + fast mode, Apache 2.0) on April 8. Mistral Medium 3 with open weights on April 9, targeting the gap between small local and large proprietary. Google's four Gemma 4 variants under Apache 2.0. Meta's Llama 4 Scout and Maverick pushing mixture-of-experts into mainstream open weights. DeepSeek R2 landed at AIME 92.7% with a ~70 percent price cut relative to frontier cloud offerings. The consensus community picks for 'run it locally on a Mac' in April 2026 are Qwen 3 32B for general reasoning and Qwen3-Coder-Next for coding.

Yes, if the local model is fronted by an Anthropic-protocol shim. Fazm's agent loop speaks the Anthropic messages API, and ACPBridge.swift lines 379 to 382 read a UserDefaults string called customApiEndpoint and inject it into the ACP subprocess environment as ANTHROPIC_BASE_URL. Point that at a local proxy that translates Anthropic request shapes into Ollama, llama.cpp, or MLX calls, and the accessibility-tree control loop runs on Qwen 3 32B or Gemma 4 without a rebuild. The SettingsPage.swift UI at lines 906 to 952 is the front door: a simple TextField with the placeholder 'https://your-proxy:8766' that the user pastes their local endpoint into.

The accessibility-tree payload inside Fazm is structured text, not pixels. That makes the text-first reasoning SKUs a better fit than the multimodal ones. Qwen 3 32B in thinking mode is a strong candidate because it was benchmarked to match or exceed GPT-4o on reasoning tasks and has stable function-calling output. Gemma 4 mid-size variants are workable if you are willing to engineer the tool-call wrapper a bit. The local-vision SKUs from April 2026 mostly underperform for this job because the control loop never needs to see a screenshot.

Three specific files: SettingsPage.swift at lines 906 to 952 renders the Custom API Endpoint card with an @AppStorage('customApiEndpoint') binding. ChatProvider.swift at line 2103 reads the same key when the bridge restarts. ACPBridge.swift at lines 379 to 382 copies the value into ANTHROPIC_BASE_URL before the ACP subprocess spawns, alongside FAZM_TOOL_TIMEOUT_SECONDS at line 387 so your local model has enough time to finish a tool call without triggering a retry. No code path in Fazm calls Ollama or llama.cpp directly; the one env var is the whole contract.

The self-sovereign framing (local, private, user-controlled weights) matches the privacy story of an accessibility-tree-based agent almost perfectly: both keep your Mac's contents on your Mac. Where they diverge is tool-call reliability. Anthropic's April 2026 Sonnet 4.6 and Opus 4.6 still have an edge on tool-format stability under chained calls, which is exactly what a Mac automation loop hammers. A local model running through an Anthropic shim in April 2026 gets you privacy and offline operation at the cost of needing a stricter tool-call validator in the shim.

Three reasons. First, the April 2026 local lineup is strong but still moving fast; pinning a weights file into a signed, notarized Mac bundle freezes the user to whatever SKU was best the week of release. Second, local model performance is tightly coupled to the user's Mac: an M1 Air and an M4 Max do not run the same model class well, and Fazm wants both users. Third, the Custom API Endpoint path at SettingsPage.swift line 936 lets any Ollama, LM Studio, or llama.cpp server the user already has running become a first-class provider without a new Fazm release.

Most screenshot-based agents need a vision-capable local SKU, which in April 2026 means big, slow, and memory-hungry. Fazm sends structured text: for each on-screen element, a role (AXButton, AXTextField, AXMenuItem), a label, a value, and coordinates, captured through AXUIElementCreateApplication and walked by a traversal in the macos-use MCP. A 7B to 32B text-first local model handles that payload fine; you do not need a 70B multimodal giant. That collapses the hardware requirement for 'run locally' by an order of magnitude.

Pick an Anthropic-protocol shim. A small proxy that accepts POST /v1/messages and rewrites it into whatever local server you are running (Ollama, LM Studio, llama-server from llama.cpp, MLX server) is ~200 lines. Run it at http://localhost:8766. Open Fazm settings, scroll to Advanced > AI Chat, toggle Custom API Endpoint on, paste http://localhost:8766, and press Enter. The SettingsPage.swift onSubmit handler at line 939 calls chatProvider?.restartBridgeForEndpointChange() which tears down the ACP subprocess and respawns it with the new ANTHROPIC_BASE_URL. Your next message goes through the local model.

Mostly tool-call format drift. Claude models produce very regular JSON inside their tool_use blocks; April 2026 local models are close but not identical, so your shim needs to validate and coerce. Second, long context. Accessibility trees for big apps (Xcode, Logic Pro) can push 100k+ tokens, and a local 32B model with a shorter context window truncates and loses tool argument fidelity. Third, the Claude Agent SDK rate-limit signals at ChatProvider.swift line 511 onward will still fire even though your local model has no rate limit, because the shim may pass through a fake 429 on backpressure.

Not yet, at least not on sustained tool-call reliability across multi-minute agent sessions driving real Mac apps. Qwen 3 32B in thinking mode is close on raw reasoning but produces noisier tool-call format under chained calls. Gemma 4 is good on instruction following but weaker on planning depth. Llama 4 Scout is surprisingly capable for its size but has quirks with nested JSON. The honest read in April 2026 is: local is viable for reflexive queries and short task loops; frontier cloud is still the pick for sustained agent work. Fazm's Custom API Endpoint lets you use both, one for the fast path and one for the hard path.