AI model releases, new papers, and open source projects in the past 24 hours (June 2026)

If you searched this expecting a tidy list of everything that shipped in the last day, here is the honest problem: a static page cannot answer a rolling 24-hour question. The list is already wrong by the time it loads. So this guide does two things. First it points you at the four sources where last-24h drops actually appear, live. Then it shows how to stop refreshing tabs and have a scheduled local agent pull that window for you every morning.

Where last-24h drops actually surface

Three primary sources cover the three things in this question, model releases, new papers, and open source projects, and each has a way to scope it to the last day. The two aggregators are useful but read-only: you cannot bend them to the topics you care about.

The catch is obvious once you do this by hand a few times: you are opening four tabs, reading the date headers, filtering out yesterday, and deduping the same release that shows up on Hugging Face and GitHub and the aggregators. That is a ten-minute chore that you will skip on the days it matters most. The fix is to make it run itself.

Turn the chore into a routine that reads your sources for you

Fazm is a native macOS app that wraps Claude Code and Codex through the Agent Client Protocol. Beyond chat, it can drive your real browser through an extension and reach native apps through macOS accessibility APIs, not screenshots. The piece that matters for this question is scheduling: the agent can create a recurring task that fires on its own, opens those four sources in your logged-in browser, and writes a deduped last-24h digest into your chat.

Because it drives your actual browser session, the routine sees the same pages you do when signed in, which matters for sources that rate-limit anonymous traffic. And because Fazm persists sessions, the routine and its output are still there after a reboot.

Building the digest in four steps

In practice you do not write any of this in a config file. You ask the agent in plain language and it calls the tool. But the underlying call looks like this, which is worth seeing because it is the part no other guide on this topic can show you:

routines_create({
  name: "Last 24h AI digest",
  schedule: "cron:0 9 * * 1-5",   // weekdays at 9am
  session_mode: "resume",          // remember what was already reported
  prompt:
    "Open huggingface.co/models?sort=created, " +
    "huggingface.co/papers, arxiv.org/list/cs.AI/recent, " +
    "and github.com/trending?since=daily. List only model " +
    "releases, papers, and repos dated in the last 24 hours. " +
    "Dedupe items that appear on more than one source. " +
    "Group by model / paper / project and keep it short.",
})

The part that makes this real: routines_create

This is not a roadmap promise. The tool is defined in the open source repo at acp-bridge/src/fazm-tools-stdio.ts (the routines_create definition sits with its siblings routines_list, routines_update, and routines_runs). Three facts from that source matter for a daily digest:

• Three schedule formats. Five-field cron (cron:0 9 * * 1-5), an interval in seconds (every:1800), or a one-shot ISO 8601 timestamp (at:2026-06-20T13:00:00Z).
• Session memory is opt-in. session_mode: "resume" keeps the ACP session across runs so consecutive digests do not repeat what they already reported. The default is a fresh session per run.
• Every run is auditable. Output lands in your chat under taskId="routine-<id>", and each execution is a row in the cron_runs table with status, tokens, cost, and timestamps, readable via routines_runs.

You can read all of it yourself. The repository is MIT-licensed and public at github.com/mediar-ai/fazm. That is the difference between this and a hosted tracker: the thing that fetches your news runs on your machine, in your browser, and you can read every line of how it decides what counts as the last 24 hours.