Latest AI model releases, papers, and open-source projects, May 29 to 30, 2026

The roundups all cover the same three feeds, and miss the layer that moved

Every write-up for a window like this points at the same places: Hugging Face sorted by created date, arXiv cs.CL recent, GitHub Trending, and the labs' own newsrooms. Those feeds are correct and worth skimming. But for a 48-hour stretch where no frontier model dropped, they also describe a quiet window, which makes it easy to conclude nothing happened. Something did. It just happened one layer down, in the application code that decides what your agent does when a key dies.

That layer almost never shows up in a release feed, because it is not a model and not a paper. It is a commit. The most honest record of what shipped on May 29 and 30, 2026 is not a list of weights; it is the diff history of the clients people actually run. Below is that record for one open-source client, pulled straight from its git log and changelog so you can verify every line.

The release that mattered: keep going on Gemini when Claude is down

The single change in this window with the largest effect on a working day is in Fazm's May 30 changelog (CHANGELOG.json, v2.9.56): "Keep going when built-in Claude is unavailable: automatically continue on Gemini if the built-in key fails or hits a usage limit, instead of showing an error." If you have ever watched an agent stop cold on a rate limit in the middle of a multi-step task, you know why this is worth more than a benchmark bump on a model you were not going to switch to anyway.

The anchor detail, the part you can check yourself: this did not arrive as one commit. It landed as a burst of nine commits on May 29 between 19:45 and 20:31 Pacific, then shipped in the v2.9.56 tag the next afternoon. The first commit is 947a5eb3 "Add state for transparent Claude-to-Gemini fallback" at 19:45:36, and ecfdbd1b "Add silent Gemini fallback on built-in model rate limits" at 20:08:21 is the one that makes the failover silent. Later commits the same evening refine it from a blunt switch into a retry queue that prefers Gemini Pro and checks availability first.

You will not find this in any release roundup for May 29 or 30. It is not on Hugging Face, it did not trend on GitHub, and no lab announced it. The only way to know it shipped is to read the commit log of a client you run. That is the recurring lesson of these dated records: the application layer is where a quiet model week is often a busy reliability week.

The dated record (primary sources)

Everything below is verifiable on github.com/mediar-ai/fazm: the dated commits on the main branch and the matching entries in CHANGELOG.json at the repo root. Timestamps are Pacific.

Two more changes worth pulling out of the log

The Gemini fallback is the headline, but two other changes in this window matter for anyone routing models or running long tasks. First, v2.9.55 on May 29 "Stopped Custom API Endpoint routing from using Fazm's built-in Anthropic key or built-in credits." Before that, pointing the app at a custom Anthropic-compatible endpoint (a corporate proxy, a GitHub Copilot gateway, or your own) could still fall back to the bundled key in some code paths. After it, a custom endpoint is fully isolated and runs on its own credentials. If you route through a gateway for compliance or cost reasons, that isolation is the difference between a clean audit and a surprise.

Second, the in-app routine scheduler that landed in commits between 20:42 and 20:55 Pacific on May 29 (RoutineScheduler, CronJobStore.dueJobs) and shipped on May 30 as "Fixed scheduled routines never running." Recurring agent jobs now fire on time while the app is open. Together with the fallback, the pattern across this window is consistent: less of the work is about which model answers, more of it is about the agent staying alive and on schedule around the model.

How to use a release the day it lands, without losing the thread

When a model or open-weight checkpoint does drop on a future day, the bottleneck is rarely finding it; it is trialing it against your real work without tearing down the session you already have open. The move is to fork the live chat. In Fazm a fork branches the current conversation into a new window carrying the full prior context, leaves the original intact and resumable, and lets you point the branch at a different backend in the same operation: Claude Code via the Agent Client Protocol, the bundled Codex (codex-acp), or a custom Anthropic-compatible endpoint.

Because there is no auto-compacting, the full history stays live for the lifetime of the window, so an A/B between two backends is measured against the same decisions and constraints rather than a silently truncated version of them. And because sessions survive a Mac restart with every window auto-restored, an evaluation that spans a reboot does not evaporate. The new Gemini fallback rides on top of all of this: if the backend you are testing rate-limits mid-comparison, the run continues instead of collapsing.

The counterargument: a quiet model week is not a slow week

It is fair to push back that a 48-hour window with no new weights is genuinely uneventful, and that pointing at one client's commit log is cherry-picking. The honest version of the claim is narrower: on any given quiet model day, the work that actually changes your throughput is happening in the application layer of tools you already run, and that work is invisible to release feeds by construction. You do not have to care about this specific repository. You should care about the method: when the roundups go quiet, read the changelog of the agent you depend on, because that is where a rate-limit failover or an endpoint-isolation fix will be hiding, and either one can matter more to your week than the next frontier checkpoint.