New AI models, papers, and open source on May 24 to 25, 2026

The honest version: the frontier took the long weekend off. Here is what actually shipped, with sources, and the argument for tracking your harness instead of the release feed, written from a Mac agent that added a whole model family to its picker the same week.

Matthew Diakonov, Written with AI

Published May 27, 20269 min

Direct answer (verified 2026-05-27)

No major lab released new model weights on May 24 or 25, 2026. May 25 was US Memorial Day, a federal holiday, and May 24 the Sunday before it, so OpenAI, Anthropic, Google, xAI, and Meta all stayed quiet. The notable open-source story was Hugging Face’s LeRobot hub passing 58,000 community datasets (reported May 25). The open-weight models people were running were from earlier in May, led by Zyphra’s Apache-2.0 ZAYA1-8B and its ZAYA1-VL-8B technical report (arXiv 2605.08560).

Check any day yourself on Hugging Face trending papers. The dates and upvote totals are stable even though the order shifts hourly.

Track the harness, not the feed

Why a quiet release weekend is the point, not the problem

A new model trends almost every day.

Most weeks, none of them change what you build.

Fazm added Gemini Flash and Pro on May 22 as a picker entry.

Your sessions, forks, and context survived the swap.

The model is interchangeable. The harness is the durable layer.

0:00 / 0:05

The dispatch: what really happened on May 24 to 25, 2026

If you typed those exact dates into a search box hoping for a single headline release, the truthful answer is that there was not one. This is not a content gap; it is the actual state of the world on a US holiday weekend. The useful thing is to separate the three layers people lump together when they search for “new model or paper or open source”: frontier weights, research papers, and community tooling. Only the third was moving.

Frontier weights

Nothing new on the 24th or 25th. May 25 was Memorial Day. The nearest frontier releases sit earlier in the month, not on these dates, so a model-release tracker shows a flat line across the long weekend.

Papers

The cleanest citation in the window is Zyphra’s ZAYA1-VL-8B Technical Report (arXiv 2605.08560), behind an Apache-2.0 vision-language MoE trained on AMD MI300 GPUs. For live trending, Hugging Face papers is the canonical record.

Community tooling

The real story. LeRobot crossed 58,000 community datasets (reported May 25), roughly 50x growth in under a year, making robotics datasets the largest single category on the Hugging Face Hub.

Open weights still circulating

Zyphra’s ZAYA1-8B (Apache-2.0, 8.4B total, 760M active) and ZAYA1-VL-8B, both downloadable on Hugging Face, were what the open community kept benchmarking over the weekend.

That is the entire honest dispatch. Three of the four boxes point at open source and tooling, and the one labeled “frontier weights” is empty. Which raises the question the rest of this page answers: if the headline-release layer is empty most days, why organize your tools around it?

The habit worth dropping: refreshing the release feed

The search behind this page is a daily ritual for a lot of engineers: open the trackers, see what dropped in the last 24 hours, decide whether to re-tool. On a Memorial Day weekend that ritual returns nothing, which feels like a dead end. It is not. It is a useful reminder that the unit you should be tracking is not the model. It is the layer that holds your work: the sessions, the forks, the context, the tool wiring. That layer should be stable enough that a new model is a small, boring change.

Here is the same week, read two ways.

Two ways to read a model-release week

Every trending model is a decision. You re-clone a setup, re-tune prompts, lose the session you had open, and start the context from zero. A quiet weekend feels like nothing happened.

New model means a migration
Sessions and context reset on every switch
A holiday weekend reads as a dead end
Your workflow is hostage to the latest drop

The anchor: Fazm’s own changelog for the same window

The one thing on this page no aggregator can copy lives in /Users/<you>/fazm/CHANGELOG.json, open-source in the Fazm repo. Three dated entries bracket the holiday weekend, and they make the thesis concrete. Two of them, on May 22, are a model arriving as a picker entry. The third, on May 25 (Memorial Day), is the kind of un-glamorous safety and recovery work that a shipping agent does regardless of what trended.

v2.9.352026-05-22model added as a dropdown
Added Google Gemini Flash (latest) and Gemini Pro (latest) as selectable models in the AI picker, alongside Claude and ChatGPT. A new model lands as a dropdown entry.
v2.9.362026-05-22model added as a dropdown
Added Google Gemini as a free option when built-in AI credits run out, next to Connect Claude and Connect ChatGPT.
v2.9.372026-05-25
Safety guardrail that warns before system-altering shell commands and prefers reversible options. Consistent handling of user passwords (never echoed, logged, or stored). Post-interrupt context widened to a 40-message history window, up from 20.

Every line above is verifiable. The repo is public, the CHANGELOG.json is committed with explicit date fields, and the signed build for each version ships on the update feed at fazm.ai/download. v2.9.35 is the load-bearing entry: “Added Google Gemini Flash (latest) and Gemini Pro (latest) as selectable models in the AI picker.” That is what absorbing a new model looks like when you build the harness instead of the model.

What stays put when you change the model

The reason a model swap is cheap in Fazm is that the model is the only thing that changes. Everything that holds your actual work is owned by the harness, not the backend.

Owned by the harness, not the model

Persistent sessions: chats survive a Mac restart, every window auto-restored with full history.
One-click fork: a new window with the full prior context, the original left untouched.
No auto-compacting: the entire conversation stays live in context for the window's lifetime.
Tool wiring: MCP servers, browser control, and accessibility-API access stay attached across a model swap.
Your account: bring your own Claude Pro or Max, or route a new model through a custom endpoint.

Fazm wraps Claude Code and Codex over the Agent Client Protocol, so the agent loop is the real thing you already use; the swappable part is which model answers. For anything Anthropic-compatible there is a custom API endpoint setting (it sets ANTHROPIC_BASE_URL under the hood), so the next open-weight model that ships a compatible gateway plugs in without an app update. The accessibility-API and screen-context tools that let the agent reach past the terminal into the browser and native Mac apps stay attached the whole time.

So what should you actually do with a quiet release weekend?

Stop treating an empty release feed as a failure of your search. Two days with no frontier weights is the normal case, not the exception, and it is the clearest argument for the setup this page describes. Keep the open trackers (Hugging Face papers is the best single one) for genuine novelty like the LeRobot milestone or a new open-weight family worth benchmarking. But put your work in a harness that does not care which model is fashionable this week: persistent sessions, one-click forks, full context that never silently compacts, and a model you pick from a dropdown. Then the next big drop, whenever it lands, is a dropdown entry, not a weekend of re-tooling.

Want to see a model swap that keeps your session, fork, and full context?

Open the Fazm Mac app with us and switch the backend mid-conversation, the same way the May 22 changelog entry added Gemini to the picker.

Frequently asked questions

What new AI models or open-source projects were released on May 24 and May 25, 2026?

No major lab (OpenAI, Anthropic, Google, xAI, Meta) shipped new foundation-model weights on either day. May 25, 2026 was US Memorial Day, a federal holiday, so the commercial frontier was quiet. The activity that weekend was in open source and community: Hugging Face's LeRobot robotics hub crossed 58,000 community-contributed datasets, a milestone reported on May 25, 2026 by TechTimes, becoming the single largest dataset category on the Hub. The open-weight releases people were still running were from earlier in the month, most notably Zyphra's Apache-2.0 ZAYA1 family (ZAYA1-8B and the ZAYA1-VL-8B vision-language model), both downloadable on Hugging Face. If you want the canonical day-by-day record, the Hugging Face trending papers page is the place to check.

Why was there no big model release on May 24 to 25, 2026?

Timing. May 25, 2026 was Memorial Day in the United States, the last Monday of May and a federal holiday, with May 24 the Sunday before it. Frontier labs almost never publish a major launch into a long holiday weekend because the announcement gets buried and the on-call coverage for a launch is thin. The frontier had also just been through a heavy April-May release cycle, so by the 24th most labs were in a build phase rather than a ship phase. The result is that anyone searching for a single headline model release on those exact dates will not find one; what they will find is steady open-source and tooling progress.

What was the biggest open-source AI story around May 24 to 25, 2026?

Hugging Face's LeRobot hub passing 58,000 community datasets. The reporting (TechTimes, May 25, 2026) framed it as an inflection point: roughly 50x growth in under a year, from about 1,145 datasets at the end of 2024 to more than 58,000 by late May 2026, making robotics datasets the largest single category on the Hugging Face Hub. The practical meaning is that any developer with an inexpensive robotic arm and a standard workstation can now fine-tune a manipulation policy against a public dataset pool large enough to sustain a self-reinforcing flywheel. The LeRobot code and datasets are open on GitHub and Hugging Face.

Were there any notable AI papers in this window?

The cleanest paper citation adjacent to these dates is Zyphra's ZAYA1-VL-8B Technical Report (arXiv 2605.08560), the report behind their open Apache-2.0 vision-language MoE (700M active, 8B total parameters), released earlier in May 2026. ZAYA1-VL-8B is built on the ZAYA1-8B base, which Zyphra trained entirely on AMD Instinct MI300 GPUs, a detail that itself drew a lot of discussion because almost all open-weight models train on NVIDIA. For a live picture of what trended on any given day, the Hugging Face trending papers list orders submissions by upvotes with stable dates.

What did Fazm ship on May 24 and May 25, 2026?

Fazm cut v2.9.37 on 2026-05-25, dated exactly that day in CHANGELOG.json in the open-source repo. It bundled the May 24 work (raising the recent-workspaces limit to 9, recording interrupted tools as cancelled, cleaning up orphaned sessions on close) plus the May 25 work: a safety guardrail that warns before running shell commands that change system state (disabling networking, killing system UI processes, rebooting) and prefers reversible options; consistent handling of user-provided passwords (never echoed, logged, or stored); and improved post-interrupt context that now carries a 40-message history window, up from 20. The more telling entry sits three days earlier: v2.9.35 on 2026-05-22 added Google Gemini Flash (latest) and Gemini Pro (latest) as selectable models in the AI picker, and v2.9.36 the same day added Gemini as a free fallback when built-in credits run out.

If a new model is released, how fast can I actually use it in Fazm?

As fast as it takes to pick it from a dropdown, because Fazm is a harness above the model rather than a model. Fazm wraps Claude Code and Codex over the Agent Client Protocol and treats the backend as swappable per chat. When Gemini became worth supporting, the change shipped as a picker entry in v2.9.35, not a rewrite. For anything that speaks an Anthropic-compatible API, Fazm also exposes a custom API endpoint setting (it sets ANTHROPIC_BASE_URL under the hood), so you can route a new model through a corporate proxy, a gateway, or a self-hosted endpoint without waiting for an app update. The point of organizing around the harness is that the model underneath can change weekly and your workflow does not.

What survives when I switch the model under a Fazm chat?

The session itself. Fazm keeps persistent sessions that survive a Mac restart, one-click chat forking that opens a new window with the full prior context while leaving the original untouched, and no auto-compacting, so the full chat history stays live in context for the lifetime of the window. Switching the backend model does not reset any of that. Contrast that with the common loop of re-cloning your setup every time a new model trends: the harness is the durable layer, the model is the interchangeable one.

How do I verify what was released on a specific day myself?

For papers and trending open models, open the Hugging Face trending papers page and read the submission dates and upvote counts, which are stable even though the ordering is volatile. For a specific project, check its GitHub releases or its model card on Hugging Face for the dated commit. For Fazm's side of this page, the source of truth is CHANGELOG.json in the public Fazm repository at github.com/m13v/fazm, where every release carries an explicit date field, and the signed builds are distributed on the update feed at fazm.ai/download.