Open-source LLM releases, April 2026, re-scored on the one input shape every leaderboard ignores

Alibaba Qwen 3 on April 8 (0.6B through 72B, dual thinking and fast modes, Apache 2.0), Mistral Medium 3 on April 9 with open weights sitting between small local and frontier proprietary, Google Gemma 4 with four variants under Apache 2.0, Meta Llama 4 Scout and Maverick pushing mixture-of-experts into mainstream open weights, and DeepSeek R2 landing at AIME 92.7% with roughly 70 percent lower inference cost than frontier cloud. All of them get benchmarked on MMLU, AIME, HumanEval, long-context retrieval, and Berkeley Function Calling Leaderboard (BFCL) numbers.

Because almost every agent framework shipping today sends screenshots to the model, not structured accessibility-tree text. The benchmarks track what the frameworks test. Fazm is one of the few consumer Mac agents that runs the control loop off real accessibility APIs through a native bundled binary (mcp-server-macos-use), so the payload the model sees is role plus label plus value plus coordinates per on-screen element, not a 1024x768 PNG. The leaderboard tells you how a model performs on a screenshot pipeline. The AX-tree payload shape is a different regime entirely.

It is a native Mach-O binary Fazm ships inside Contents/MacOS/mcp-server-macos-use of the signed, notarized app bundle. The ACP bridge references it at acp-bridge/src/index.ts line 63 as macosUseBinary = join(contentsDir, 'MacOS', 'mcp-server-macos-use') and registers it as the 'macos-use' MCP server at line 1059 with no Python and no Node wrapper. Every AX tool call an LLM fires inside Fazm runs through that binary and out to AXUIElementCreateApplication on the frontmost process.

The text-first reasoning SKUs win because the payload is structured text. Qwen 3 32B in thinking mode is the strongest practical local pick: it handles deep nested tool-call JSON and the accessibility tree's long-tail of sibling elements. DeepSeek R2 is also very capable if you are willing to run it through an API gateway rather than locally. Gemma 4 mid-size works for short task loops. Llama 4 Scout is surprisingly capable for its footprint but gets quirky with deeply nested JSON tool arguments, which matters once an Xcode or Logic Pro AX tree lands in the context.

Because the agent loop never needs a screenshot. The accessibility tree carries the role, label, value, and frame of every element the user can interact with. A vision SKU spends memory and parameters on a visual encoder that never runs. In April 2026 that translates directly into hardware cost: a 70B multimodal model that a user cannot load on a 16GB M-series Mac loses to a 32B text-first model that fits, and the text-first model has equal or better structured output stability.

Through an Anthropic-protocol shim. Fazm's agent speaks the Anthropic messages API, and ACPBridge.swift lines 379 to 382 copy a UserDefaults string called customApiEndpoint into the ACP subprocess environment as ANTHROPIC_BASE_URL. Point that endpoint at a proxy that rewrites Anthropic requests into Qwen 3, Gemma 4, or Mistral Medium 3 calls, and every tool_use block the model emits reaches the macos-use MCP server the same way a Claude Sonnet 4.6 tool call would. No rebuild, no fork, no code change.

Each element is a single text row with a role (AXButton, AXTextField, AXMenuItem, AXStaticText, AXCheckBox), a human-readable label, a value, x and y coordinates, width and height, and a visibility flag. The traversal is produced by AXUIElementCopyAttributeValue calls in the bundled binary and flattened into readable rows so a text-first LLM can select an element by label and role without parsing pixels. The shape is deterministic and stable across apps, which is exactly the kind of input April 2026 reasoning models handle cleanly.

BFCL grades whether a model produces syntactically valid JSON against curated schemas and semantically correct arguments against toy functions. It does not exercise long, live, repeatedly refreshed accessibility trees with hundreds of sibling nodes, concurrent apps, and coordinate-aware click arguments. A model that scores 90+ on BFCL can still stall on a real AX payload when the tree is 15K tokens, the user has three Slack windows open, and the correct target is the second AXButton labeled 'Send' in the third window. The right benchmark for this job would measure end-to-end task completion against a real app, not schema conformance.

Not yet on sustained tool-call reliability across multi-minute agent runs. Qwen 3 32B in thinking mode is the closest practical match on raw reasoning. Gemma 4 is strong on instruction following but weaker on deep planning. Llama 4 Scout is surprisingly capable. DeepSeek R2 rivals frontier on reasoning. The honest read in April 2026 is that local and open-weights April models are viable for reflexive single-step queries and short agent loops, and frontier cloud is still the pick for sustained end-to-end Mac automation. Fazm's Custom API Endpoint lets the user run both: a cheap fast path and a strong hard path, against the same AX-tree payload.

Start with the input shape, not the leaderboard. If you send screenshots, compare vision SKUs on pixel-grounding latency and OCR fidelity. If you send structured accessibility-tree text (what Fazm does), favor text-first reasoning SKUs that handle 10K to 30K tokens of nested structured output reliably. Qwen 3 32B is the default pick, with Gemma 4 mid-size as a memory-constrained fallback and DeepSeek R2 through an API gateway when you want frontier reasoning without the frontier price.