Hugging Face, April 2026, from a Mac desktop agent viewpoint

The April 2026 drops leaned multimodal and OCR-heavy. Google shipped the Gemma 4 family (E2B, E4B, 26B MoE, 31B dense, all Apache 2.0) with vision. Mistral added Mistral 4 with vision and Codestral 2 (22B, Apache 2.0). Baidu pushed PP-OCRv5 and a document-understanding model called UVDoc. Cohere landed its first speech-recognition models on the Hub. Netflix published a video-inpainting model. Alibaba continued filling in Qwen 3 variants. And on the audio-reasoning side, NVIDIA released Music Flamingo plus the NomicBERT v3 embeddings.

The most important Hub change in April was not a dataset, it was a feature: Datasets now natively ingests agent traces (sessions, turns, tool calls, model responses) with a dedicated viewer. Before April, sharing agent traces meant coaxing JSONL into the Datasets schema by hand. Now uploading a run of a desktop agent produces something browsable on the Hub. Alongside that, the CVPR 2026 Foundation Models for General CT Image Diagnosis datasets shipped, and the open-gigaai/CVPR-2026-WorldModel-Track-Dataset went live for video world-model research.

No. On macOS, any app that implements accessibility exposes a structured tree through AXUIElement. Fazm reads that tree directly. AppState.swift line 439 calls AXUIElementCreateApplication(frontApp.processIdentifier), then line 441 calls AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute as CFString, &focusedWindow). The result is typed data (roles, titles, values, positions) returned in microseconds with no model in the loop. You would only reach for PP-OCRv5 or UVDoc when the surface is pixels (a Qt app, a Flutter canvas, an OpenGL viewport, a shared screen in Zoom) where the accessibility tree is empty.

Raw LLM traces record prompts and completions. Agent traces record the full loop: the session, each turn, the tools the model called, the arguments it passed, and the results that came back. For a desktop agent, that includes every click, type, and accessibility query. The Hub's April 2026 upload flow auto-detects the agent-trace schema and renders a session browser instead of a flat JSONL view. That matters because desktop-agent failures are almost always sequence failures, not single-turn failures.

Not yet for full desktop automation. Gemma 4 E2B and Gemma 3n (2B effective footprint) are close enough to run on an M-series Mac, and Mistral 4 fits on a 48GB MacBook Pro. But sustained structured tool-calling over a twenty-step accessibility-tree workflow still favors hosted frontier models. Fazm's acp-bridge/src/index.ts defaults every new session to claude-sonnet-4-6 because that is the cheapest model with the tool-call throughput the loop needs. The gap is closing, but it closes from the tool-reliability side, not the benchmark side.

In principle yes, and the new HF agent-trace upload flow makes the ingestion half trivial. In practice the blocker is licensing and data provenance. Accessibility-tree snapshots contain user data (app names, window titles, sometimes document contents). Any dataset you upload needs either synthetic traces or explicit user consent, and the target model's license has to permit commercial fine-tuning. Apache 2.0 (Gemma 4, Qwen 3, Codestral 2, OLMo 2) and MIT (GLM-5.1) are safe; the Meta Community License for Llama 4 has a 700M MAU cap; the Gemma license is Google-specific.

Because almost no one is training agents on accessibility trees. The dominant paradigm is still vision-first: train on screenshots plus synthetic action labels. That is why the April drops included PP-OCRv5 and UVDoc but nothing native-tree-shaped. The irony is that accessibility trees are cleaner, smaller, and higher-signal than screenshots for any platform that exposes them (macOS, Windows, iOS, Android). The opportunity for a community dataset here is real, and the new HF trace-upload feature is the first piece of infrastructure that makes it possible without custom tooling.