The latest AI news on Hugging Face and GitHub, cross-referenced by a Mac agent that reads both at once

April 2026 was dense. On the model side: Codestral 2 from Mistral, Qwen3-Coder-32B and smaller Qwen3 coder variants from Alibaba, Meta's Llama 4 Maverick and the llama-stack repo cluster around it, Google's Gemini-based open weights paired with the Agent Development Kit (google/adk-python), Ai2's OLMo refresh, and a stream of small-VLM releases including SmolVLM2-2.2B. Hugging Face itself acquired Pollen Robotics in spring 2026 and wrapped its LeRobot repo into the Hub so datasets and model cards live alongside code. The pattern: almost every release lands as a model card on huggingface.co and as a repo or release on github.com on the same day, often with the GitHub README linking to the HF card and vice versa. That is the cross-reference worth automating.

Because trending is not release. huggingface.co/models?sort=trending is vote-weighted and lagging. A model that a major lab dropped five hours ago may not be on trending yet, and a model that trended last week still sits there because of inertia. huggingface.co/models?sort=downloads is a popularity list, not a news feed. huggingface.co/papers/trending is weighted toward arXiv preprints, not actual weights you can run. The release signal you want is huggingface.co/models?sort=created&direction=-1, which is sorted by upload time and shows exactly what shipped in the last N hours. Nobody links to that URL because it does not have a nice name; it is the right URL anyway.

Because Hugging Face model names are full of non-Latin characters. tencent/HY-OmniWeaving, alibaba/Qwen3-Coder-32B, ByteDance/Doubao-1.5, Stability AI/japanese-stable-diffusion-xl, and a long tail of community repos named in Chinese, Japanese, Korean, Arabic, or with math symbols. OCR error rates on these names are high even for SOTA VLMs. When Fazm reads the Hugging Face model list through the macOS accessibility tree it gets model names back as actual UTF-8 strings, not pixel guesses. The specific call is AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute as CFString, &focusedWindow) in Desktop/Sources/AppState.swift line 441. Every card in the results grid becomes a typed AXGroup with labeled children for name, base model, task, license, and download count. The LLM sees strings, not screenshots.

You need two, one per platform, because the two sites expose different sort keys. For Hugging Face: https://huggingface.co/models?sort=created&direction=-1 filtered by pipeline_tag=text-generation (or image-text-to-text, feature-extraction, etc.) with a date window. For GitHub: https://github.com/search?q=topic%3Aartificial-intelligence+pushed%3A%3E2026-03-31&s=updated&type=repositories covers the whole month. A Mac agent runs both in one pass and joins on model name or org handle. The full-text join is usually enough because labs use the same repo name on both sites, e.g. Qwen/Qwen3-Coder-32B on HF matches QwenLM/Qwen3-Coder on GitHub.

In a local SQLite database at ~/Library/Application Support/Fazm/fazm.db. The execute_sql tool declared at acp-bridge/src/fazm-tools-stdio.ts line 281 lets the agent INSERT into any table you define. A reasonable schema is ai_release_cross_ref(date, name, hf_url, gh_url, hf_downloads, gh_stars, license, task, notes). The table is append-only in practice; each day's run gets a new date and you can SELECT * FROM ai_release_cross_ref WHERE date = date('now') to see what dropped today. SELECT auto-limits to 200 rows, UPDATE and DELETE require WHERE, DDL is blocked. That is the full contract.

Yes. The repo is github.com/mediar-ai/fazm. License is MIT. The two lines that matter for this workflow are in Desktop/Sources/AppState.swift at lines 439 to 441: AXUIElementCreateApplication(frontApp.processIdentifier) followed by AXUIElementCopyAttributeValue(appElement, kAXFocusedWindowAttribute as CFString, &focusedWindow). You can grep for them, run the app with ./run.sh, and inspect every AX call it makes against the Hugging Face or GitHub window.

Those are editorial. The HF newsletter summarizes what the HF team found notable; the state-of-open-source posts run on a quarterly cadence. Neither answers the question a working engineer asks at 9am: what did my lab, my competitor, or my favorite open-weights team ship in the last 24 hours, and where is the code. The cross-reference digest does. It takes the HF created-sort list and the GitHub pushed:> list, intersects them by name or author, and hands you a table you can query. The editorial feeds stay useful for narrative; the local digest is the raw signal.

No. Fazm is a consumer Mac app. Install the DMG, grant Accessibility and Screen Recording permission during onboarding, open the chat, and say something like 'every morning at 8am, open huggingface.co/models?sort=created&direction=-1 and github.com/search?q=topic:artificial-intelligence+pushed:>yesterday, read the top 30 results from each, join them by name, and write the result to a table called ai_release_cross_ref.' The agent calls capture_screenshot, web-scraping, and execute_sql on your behalf. The Swift and Rust references in this article are for the subset of readers who want to audit what the app actually does.

There are none in this workflow. Fazm reads the rendered HTML of the Hugging Face and GitHub pages through the accessibility tree of a visible Chrome or Safari window. No huggingface.co API token. No GitHub personal access token. No SaaS quota. The only thing the agent needs is an open browser tab on the right URL and Accessibility permission for Chrome (or Safari) in System Settings. If you already use the Hugging Face Hub API for downloads, that runs separately and stays unaffected.

Desktop/Sources/BundledSkills/web-scraping.skill.md. It lives alongside sixteen other bundled skills (deep-research.skill.md, pdf.skill.md, docx.skill.md, video-edit.skill.md, and more). The web-scraping skill tells the agent to fall back to requests + BeautifulSoup for static HTML or Playwright for JS-rendered pages. Fazm ships its own bundled Python at Desktop/Sources/Resources and references that path via the bundled_python_path constant in Desktop/Sources/Chat/ChatPrompts.swift. So even on a Mac where the user has never installed Python or pip, the fallback works.