The weights got loud. The consumer agents stayed quiet.

The month kept the pattern that has held for most of 2025 and 2026: Meta shipped refresh builds of the Llama family, Alibaba shipped a new Qwen3 checkpoint, DeepSeek shipped incremental V3 weights, Mistral shipped a smaller reasoner tuned for local inference, and Google pushed a fresh Gemma drop tuned for on-device. The specific version strings matter less than the shape: every major lab now treats open weights as a standing release track, and the benchmark gap between the best open weight and the best closed model keeps tightening. The actual story, though, is not on the weights side. It is that a non-developer consumer still has very few ways to hold any of these models day zero in a product they would pay for.

Fazm is not a model. Fazm is a MIT-licensed consumer Mac agent. The model it talks to today is Claude, through the Claude agent SDK bundled as an ACP bridge subprocess. The relevant open source story is on the client side: the Swift app source, the bridge source, the skills, the installer, all of it lives on GitHub under an MIT license. That is the layer most of April 2026's open source LLM coverage skips. A list of open weight models is not a product. A product that can hold them is.

Not today, honestly. The whole model list pipeline is bound to one bridge subprocess that speaks ACP and talks to the Claude agent SDK. What exists today is the seam: emitModelsIfChanged in acp-bridge/src/index.ts at line 1271, plus the 4-row modelFamilyMap in Desktop/Sources/FloatingControlBar/ShortcutSettings.swift at line 159. A second bridge subprocess that spoke the same IPC shape against a local llama.cpp or MLX runtime would flow through the same code. The picker already tolerates unknown model IDs: the else branch at ShortcutSettings.swift line 189 assigns sort order 99 to any ID that does not contain haiku, sonnet, opus, or default. So the day a second bridge existed, an open weight model ID would show up at the bottom of the picker the same launch.

A model on HuggingFace is a download, not a workflow. Getting from a 40 GB weight file to a grandmother-uses-it agent requires: a runtime that can serve the model on consumer hardware, a daemon that stays alive, a UI that lets a non-developer pick the model, an accessibility layer that lets the agent see and act on a real screen, and an update channel that can keep the whole stack current. The model is one of six layers. In April 2026 the model layer has five new entries this month. The other five layers have roughly zero new MIT-licensed, consumer-ready entries this month. That is the gap this piece is about.

It lives in Desktop/Sources/FloatingControlBar/ShortcutSettings.swift at line 159 and has exactly four rows: (haiku, Scary, Haiku, 0), (sonnet, Fast, Sonnet, 1), (opus, Smart, Opus, 2), (default, Smart, Opus, 2). Each row is (substring, shortLabel, familyLabel, sortOrder). When a new model ID arrives from the ACP bridge, updateModels at line 180 substring-matches it against the table. A match yields a human label like 'Fast (Sonnet, latest)' and a known sort order. A miss lands at the else branch at line 189, which uses the raw API name and assigns sort order 99. Everything in the picker, from the floating bar to the Settings panel, reads off the same availableModels array.

Because the map is substring-matched, not full-string-matched. A row like ('llama', 'Local', 'Llama', 3) would cover llama-3, llama-4, llama-4.1-instruct, llama-4-70b-mlx, and any future naming variant the community settles on, in one line. The picker would put it at sort order 3, right after Smart. The relabel would read 'Local (Llama, latest)'. Qwen, DeepSeek, Gemma, Mistral each become another single row. Anthropic's 4 rows expanding to 9 would give the picker coverage of every major open weight family without a single hard-coded version string.

Because the problem it solves is very small. Take a list, drop one entry with a known pseudo-ID, JSON-stringify it, compare against the memoized previous string, fire one IPC message if different. Everything else (the UI rebuild, the picker order, the keyboard shortcut list) is downstream of that single 'models_available' message. Wrapping this in a class would add surface without removing anything. The 10-line function also caps context cost for every LLM that later reads it, which is a real consideration when your own product is driven by one.

The repo is at github.com/m13v/fazm, MIT-licensed. Three files matter: Desktop/Sources/FloatingControlBar/ShortcutSettings.swift for the 4-row modelFamilyMap (line 159) and the updateModels sorter (line 180), acp-bridge/src/index.ts for emitModelsIfChanged (line 1271) and its two call sites at lines 1347 and 1361, and README.md for the MIT license line and the 'Fully open source. Fully local.' tagline. Clone, grep, read. Nothing in this piece is abstract: every line number resolves to code you can hold.

Locally, there is nothing blocking it architecturally. The same IPC shape ('models_available' carrying a list of modelId, name, description) works equally well for a local runtime. The real constraints are device, not software: a useful open weight agent model wants RAM and a fast GPU, and the experience is very different on a 16 GB Air versus a 64 GB Studio. The honest answer is that the bridge design already supports it, and the product team has not yet committed to a specific open weight model because the consumer-surface work is the larger cost, not the inference call.