Hugging Face new models, April 2026: the client-side hookup nobody writes about

Google's Gemma 4 family landed first, with 1B, 13B, and 27B multimodal variants plus experimental any-to-any E-series weights. Alibaba's Qwen 3.5 followed with the 9B base topping downloads at 4.8M copies pulled in the first window, and community fine-tunes from Jackrong and HauhauCS claiming three of the top five like counts. MiniMax shipped M2.7, Mistral shipped Codestral 2 22B and Mistral 4, Meta's Llama 4 Scout 17B and Maverick 17B went MoE, DeepSeek V3 Base refreshed, Netflix entered with a video inpainting model, Cohere extended into speech recognition, and Jina shipped embeddings v3. The shared architectural thread is MoE going mainstream: 70B-class quality at 13B-class dense inference cost.

You front the local model with an Anthropic-shape shim (vLLM, Ollama, or a proxy in front of either) and point Fazm's Custom API Endpoint at that URL. That single field rewires every tool-using chat in the app to your local endpoint. The entire integration surface on Fazm's side is three lines in Desktop/Sources/Chat/ACPBridge.swift lines 378 to 382, which read the customApiEndpoint UserDefaults key and inject it as ANTHROPIC_BASE_URL before the ACP subprocess spawns. No per-model code, no special adapter, no vllm.swift, no ollama.swift.

Text-only works. Fazm does not send screenshots. It captures macOS accessibility trees directly via AXUIElementCreateApplication, starting around line 439 of Desktop/Sources/AppState.swift. The payload that lands on the inference endpoint is structured text: role, label, value, and coordinates for each on-screen element. That means Qwen 3.5 9B base, Gemma 4 1B and 13B base, DeepSeek V3 Base, Llama 4 Scout 17B, and Mistral 4 text variants are all viable desktop-agent backbones. You can skip the multimodal SKUs and the VRAM overhead that comes with them.

On a 64 GB M3 Max, quantized Qwen 3.5 9B in 4-bit comfortably fits with room for a 16k context window. Gemma 4 1B runs on basically anything. Gemma 4 13B quantized is workable. SmolVLM2 2.2B from HuggingFaceTB (vision, if you want it) runs fine. Llama 4 Scout 17B MoE has an effective dense footprint around 3.5B active parameters per token and quantizes down to a 5-6 GB inference budget. MiniMax M2.7 is designed to be svelte. On a 128 GB M3 Ultra Mac Studio, Gemma 4 27B and Qwen 3.5 32B fit with headroom. vLLM on Apple Silicon has improved through 2026 but llama.cpp and MLX are still the smoother Mac-native paths.

Six steps. 1) On a Linux box or Mac Studio, vllm serve google/gemma-4-13b --host 0.0.0.0. 2) Put an Anthropic-shape shim in front of it (there are several open-source options, all POST /v1/messages shaped). 3) Verify the shim with curl -X POST https://your-shim/v1/messages -d '{...}' and check you get back a well-formed assistant turn. 4) In Fazm, open Settings, Advanced, Custom API Endpoint, paste the shim URL, hit Enter. That .onSubmit fires restartBridgeForEndpointChange in ChatProvider.swift. 5) Start a new chat and send something like 'summarize my active Safari tab.' That exercises accessibility-tree capture, the bridge, the shim, and Gemma 4 end-to-end. 6) If it roundtrips, you have a local Hugging Face model driving a Mac desktop agent.

Because screenshot agents force you to use vision-capable models, which for the April 2026 cohort means Gemma 4 multimodal variants, Llama 4 multimodal paths, or SmolVLM2. Those are strictly larger and more expensive than their text-only siblings at equivalent reasoning quality. When your agent sends a structured accessibility tree instead, the model only needs to be good at reading text and choosing actions. Qwen 3.5 9B base is excellent at that. You get production-adjacent desktop automation on a single 24 GB GPU, or a quantized 9B on a recent Mac, without paying the multimodal tax. The April 2026 Hugging Face wave is disproportionately useful to accessibility-tree agents for this reason.

No, and that is deliberate. Fazm ships with Claude Sonnet 4.6 as the default backbone. The Hugging Face integration is a user-configurable URL in the Custom API Endpoint field, which maps to a UserDefaults key that gets injected as ANTHROPIC_BASE_URL at ACPBridge.swift line 381. There is no vllm adapter, no ollama adapter, no llama.cpp branch in the codebase. That is the entire story, and it means every new April 2026 Hugging Face model works the day it lands on the Hub, assuming your shim does.

Three things, in order of how often you hit them. Tool-use format stability: 9B models follow Anthropic tool-call schemas less reliably than Claude Sonnet 4.6, so you get occasional malformed JSON that the ACP bridge has to tolerate. Multi-step planning depth: complex chains involving five or more apps degrade earlier. Ambiguous accessibility nodes: when the tree has two buttons labelled 'Send,' a 9B model is more likely to pick the wrong one. These are tractable with tighter system prompts and a retry policy, but they are the real-world gap between a 9B local model and the hosted frontier. None of them are about raw perception, which is exactly the point: the accessibility tree removed the perception problem.

Three tasks, each takes under a minute end-to-end. Task A: 'Summarize my active Safari tab in three bullets.' Exercises tree capture, content extraction, summarization. Task B: 'Reply to the most recent email in Mail with a one-line thank-you.' Exercises multi-step navigation, typing into a specific field, choosing the right button. Task C: 'Find the file named notes.md in Finder and move it to Desktop.' Exercises file-operation planning and physical-looking navigation without any pixels. Run each on Claude Sonnet 4.6 first to establish a success baseline, then swap the endpoint to your Hugging Face backend and rerun. The delta tells you exactly where a given model sits.