Open-source LLM releases, April 2026: one env var swaps any of them into a shipping Mac agent

The short list is DeepSeek-V3.2 (MIT, strong reasoning and tool-use), Qwen 3.5 (Apache, 122B total parameters with 10B active via MoE, runs on a 64 GB MacBook), Meta Llama 4 Maverick and Scout (native multimodal, MoE), Google Gemma 4 (26B MoE flagship, about 14 GB on disk, roughly 85 tokens per second on consumer hardware), and Mistral Large 3 (the company's EU-residency-positioned flagship). Most recap pages stop at the benchmark table. The question this guide covers is what any of those releases actually looks like when you try to pipe it into a shipping consumer Mac app.

Yes, as long as you put an Anthropic-shape messages API in front of it. Fazm's main control loop reaches the model through the ACP bridge, which reads the customApiEndpoint UserDefault (see Desktop/Sources/MainWindow/Pages/SettingsPage.swift line 840) and injects it as ANTHROPIC_BASE_URL at Desktop/Sources/Chat/ACPBridge.swift line 381. Point that URL at a local proxy that translates Anthropic messages requests into whatever protocol your open-source model speaks (Ollama, vLLM, TGI, LM Studio, llama.cpp), and the rest of the app (tool calls, the macos-use MCP server, the accessibility tree capture) keeps working unchanged.

Most other desktop agents send screenshots to their LLM, so they are forced to run on a multimodal model. That rules out most April 2026 open-source releases unless you pick Llama 4 Maverick or another native-multimodal SKU. Fazm sends the macOS accessibility tree as structured text (role, label, value, coordinates for each element, captured through AXUIElementCreateApplication). The model never sees a pixel. That means any strong text-and-tool-use open-source model, including DeepSeek-V3.2 and Qwen 3.5 which are text-first, becomes a first-class candidate. The most hyped multimodal benchmarks on the April 2026 open-source charts are not the relevant axis for this kind of agent.

At minimum it needs to serve POST /v1/messages, accept the Anthropic message format (system prompt, messages array with role/content, tools array with JSON Schema input_schema, tool_use and tool_result content blocks), and return either a final message or a stream of events that the Claude Code runtime understands. The shim translates that into whatever the underlying open-source model speaks (OpenAI-shape JSON for Ollama, Qwen's native format, vLLM's OpenAI-compatible endpoint, etc.). Projects like Ollama with the claude-relay pattern, litellm's anthropic adapter, and any of the one-file Python proxies on GitHub are enough for a weekend experiment. Production use is a different story because Fazm's agent loop fires many tool calls per minute and shim failures surface as stuck sessions.

No. ShortcutSettings.defaultModels (Desktop/Sources/FloatingControlBar/ShortcutSettings.swift lines 151-155) hardcodes the three Claude tiers: Haiku, Sonnet, Opus. Gemini 2.5 Pro is wired into a separate passive screen-observer loop (GeminiAnalysisService.swift line 67: private let model = 'gemini-pro-latest'). Any April 2026 open-source model gets in through the Custom API Endpoint field in Settings, which the SettingsPage UI labels as an advanced setting for proxies, corporate gateways, and custom shims. The moment you flip the toggle on and paste your proxy URL, Fazm restarts the bridge (ChatProvider.restartBridgeForEndpointChange) and the next query runs against whatever open-source model sits behind that URL.

DeepSeek-V3.2 is the current working answer, for three concrete reasons. First, its tool-use format stability is the best of the open-source pack, which matters more than raw reasoning score when the model spends its life emitting tool_use blocks with structured JSON arguments. Second, it is released under MIT, so a self-hosted shim is legally simple for commercial use. Third, its token cost on a hosted inference provider (Together, Fireworks, DeepInfra) is roughly an order of magnitude below frontier Claude, which is the real-world rate-limit problem Fazm hit in April 2026 with Opus 4.6 and now with agent loops generally. Qwen 3.5 is second, mainly because at 10B active parameters it runs on a 64 GB MacBook locally, which matters for the privacy-first user segment.

Three things, in order of annoyance. Tool-use format drift: open-source models occasionally emit malformed tool_use arguments, and the Claude Code runtime rejects those as non-retryable, which shows up as a stuck session in the ACPBridge stderr log. Context-window mismatches: Claude Sonnet 4.6 expects a 200k context window for long sessions, and if your shim silently truncates to the open-source model's limit, the agent starts forgetting mid-task. Streaming event shape differences: Anthropic-shape streaming uses a specific sequence of content_block_delta events. Shims that approximate this break the UI's typing indicator. None of these are fatal, but all three are why open-source models are not currently Fazm's default.

Yes, with the right pair of tools. Run Qwen 3.5 via LM Studio or Ollama locally, put a shim like claude-relay or litellm (configured with the anthropic adapter) in front of it on localhost:8766, and paste http://localhost:8766 into Fazm's Custom API Endpoint field. The main agent loop now runs fully offline against Qwen. The separate Gemini screen-observer loop still reaches out to Google for multimodal analysis, but that loop is opt-in through Session Recording and can be left disabled. On a 64 GB M-series MacBook, Qwen 3.5 at 10B active parameters plus Fazm's accessibility-tree input keeps the agent loop interactive for the kinds of tasks where you previously used Sonnet.

Not by default. Fazm's main agent loop does not send images to the model, so Llama 4 Maverick's multimodal capability is unused in that loop. If you wanted Llama 4 to drive a screen-observer loop comparable to the Gemini 2.5 Pro one in GeminiAnalysisService.swift, you would need to build that loop yourself, because the screen-observer path does not flow through ANTHROPIC_BASE_URL, it calls Google's generative AI SDK directly. For the text-only control loop, picking Llama 4 over Qwen 3.5 or DeepSeek-V3.2 is a tradeoff between token cost, tool-use format reliability, and licensing terms, not a multimodal question.

Because everything on this page is traced to a file and a line in Fazm's shipping desktop codebase, not to a press release. The Custom API Endpoint toggle is at SettingsPage.swift line 840 with the UI at lines 906-950. The env-var injection that routes every Claude-shaped request through the proxy is one line, ACPBridge.swift line 381. The default model list that the picker ships with is ShortcutSettings.swift lines 151-155. The passive screen observer that stays on Gemini and does not touch the open-source path is GeminiAnalysisService.swift line 67. Those references are the difference between what the April 2026 leaderboards say and what a real product actually does with those releases.