Local LLM releases, April 2026: a month of weights, and the macOS half nobody writes up

Between April 2 and April 14, six organizations pushed open-weight releases. Google Gemma 4 landed on April 2 under Apache 2.0 in four deployment-targeted sizes. Alibaba Qwen 3 shipped on April 8 across 0.6B, 1.7B, 4B, 8B, 14B, 32B, and 72B, with a dual-mode system where the same weights run in a reasoning mode and a standard mode. Mistral Medium 3 followed on April 9 with open weights aimed at the gap between small local models and frontier proprietary ones, priced at $2 per million input and $6 output on Mistral's hosted API for reference. Meta Llama 4 ship dates drifted across the month, with the Scout 17B Mixture-of-Experts variant leading the Ollama downloads for 12GB VRAM machines, and the Maverick variant stretching a 10M token context window on higher-end rigs. These are the four families that dominated April downloads on Ollama, LM Studio, and Hugging Face.

Because a language model produces text. An agent produces clicks, keystrokes, file saves, email sends, and API calls. The bridge between the two is what most guides skip. On macOS specifically, the bridge is the system accessibility layer: NSAccessibility and AXUIElement, the same APIs that power VoiceOver. Every native Mac app exposes its UI as a tree of AXUIElement nodes with role, label, value, and position. A model can read that tree, decide what to do, and post a synthetic AXPress or AXSetValue back. That is the shortest, most reliable loop between weights and action on a Mac. You can skip it and send screenshots into a vision model, but you will pay for it in latency, tokens, and brittleness every time a modal pops up or a sidebar changes width.

In acp-bridge/src/index.ts at line 63, the file resolves a path to a native binary called mcp-server-macos-use at Contents/MacOS/mcp-server-macos-use inside the app bundle. At line 1056 the bridge checks that the binary exists, then registers it as an MCP server named macos-use at line 1059 with no arguments and an empty env. That server is one of exactly five built-in MCP names listed at line 1266: fazm_tools, playwright, macos-use, whatsapp, google-workspace. When a Claude or agent query needs to read or act on a Mac app that is not the browser, the call goes through the macos-use MCP, which is doing AXUIElement traversal, not image decoding. That is the behavior that makes the product work with any app.

Today Fazm runs the orchestration through Claude via the ACP bridge. The macos-use binary itself is protocol-plain MCP, which means anything that speaks Model Context Protocol over stdio can consume it. If you stand up an MCP-capable local runtime, say a Qwen 3 32B or Llama 4 Scout served through an MCP client shim, you can register the same binary as a server there. The binary does not care what is on the other end of the pipe, only that the caller speaks MCP. So the accessibility half of the story is model-agnostic even though the consumer app ships with Claude.

Four reasons. First, determinism: AXUIElement returns structured data with stable identifiers, so 'click the Save button in Figma' is a deterministic tree walk, not a pixel-matching gamble. Second, permissions: a screenshot pipeline on macOS requires Screen Recording and triggers the recording indicator at the menu bar the entire time the agent is active. Accessibility does not. Third, latency and tokens: an accessibility snapshot of a focused window is kilobytes of text, a screenshot is hundreds of kilobytes of base64 that the model has to reason over. Fourth, occlusion and scale: screenshots fail at arbitrary zoom, multiple monitors, and full-screen toggles. Accessibility trees survive all of that because they are not pixel-space.

For a 24GB Mac with unified memory, Qwen 3 32B at 4-bit quantization is the strongest point on the curve for tool-use and reasoning, and it ships with a dual-mode that lets you pay for thinking only when the task needs it. For a 16GB machine, Llama 4 Scout 17B MoE runs well because the active-parameter count is low even though the total parameter count is higher. For a 64GB M-series machine, Mistral Medium 3 is the most capable local option for mixed-language, document-heavy work. None of these ship with native MCP client behavior, so to plug any of them into a desktop automation loop you need a client shim on top that translates their tool-call outputs into MCP requests.

Not today. Fazm's consumer product runs Claude by default because the reliability of the full agent loop (tool selection, argument shaping, recovery from errors, multi-step planning) is currently set by frontier model quality, and Opus 4.7 was a meaningful jump there on April 22, 2026. The local side of the story is that the rest of Fazm is already local: the app, the database, the knowledge graph, the browser-profile extraction, and most importantly the accessibility automation itself all run on-device. Swapping the brain to a local Qwen 3 or Llama 4 is a future lever, not a present default. The open-source repo at github.com/mediar-ai/fazm is what makes that lever pullable.

Two of the four families went MoE. Llama 4 Scout and Maverick are both Mixture-of-Experts, which is why Scout at 17B active parameters runs on a 12GB VRAM card despite having a much larger total parameter count. Qwen 3 stayed dense across the family, which makes it easier to quantize cleanly at 4-bit and 8-bit without the expert-routing artifacts you sometimes see. Mistral Medium 3 is dense. Gemma 4 is dense. The practical takeaway: if you have VRAM constraints and want the strongest parameter count per GB, Llama 4 Scout is the answer. If you want the most predictable quantization behavior and the widest quant-kernel support across llama.cpp, MLX, and vLLM, the dense Qwen 3 family is the safer pick.

MCP, Model Context Protocol, is an open stdio-based protocol for servers that expose tools to an LLM client. Each server ships a tool schema, the client calls tools, the server returns results. The reason MCP matters specifically for this month's releases is that every April 2026 open model is good enough at tool-use that the bottleneck stops being the model and starts being the ecosystem of tools it can actually reach. macos-use is one such server. Playwright, whatsapp, and google-workspace are three more that Fazm bundles. The moment a local model speaks MCP, everything built for frontier-model tool-use ecosystems becomes available to it without a rewrite.

Claude Opus 4.7 went GA with a 1M context window and a material jump in long-horizon coding quality. That is not a local release, but it resets the ceiling that local models are chasing, and in practice it is the reference point against which the April open-weight releases are measured on real agent tasks. The other thing that shipped the same weekend was a refreshed Claude consumer terms page, which broke a lot of client apps that did not distinguish a terms-not-accepted 400 from an auth-failure 400. That second change is invisible to a user running a fully local loop, which is one of the arguments for running a fully local loop in the first place.