New AI model releases, papers, and open source around May 29, 2026, and how to actually use them

The feed is the source of truth, not the roundup

If you searched for the releases on a specific late-May day, you already know the frustration: most write-ups are out of date before they finish loading. Open-weight families such as Llama, Qwen, DeepSeek, Mistral, and Gemma push new checkpoints and fine-tunes constantly, and the language-and-computation preprint feed runs dozens of papers a day. A static list is a photograph of a river.

So the first move is to stop treating any article as the registry. Bookmark the three live feeds above and skim them on the cadence you care about. The three sources below are the ones that stay correct without anyone maintaining them.

The headline is the model. The throughput is the harness.

Here is the uncomfortable part for anyone refreshing a release feed: for everyday coding and desktop work, swapping one frontier checkpoint for another rarely changes your output as much as the loop you run it in. Four properties tend to decide whether a new release actually pays off, and none of them live in the weights.

This is why a daily list, on its own, is thin advice. Knowing that a model exists is the easy 5 percent. The other 95 percent is whether your tooling lets you fold it in without losing state, context, or reach.

How to test a release in minutes, without losing your work

The fastest honest test is a side-by-side on a real task. The catch is that most setups make you tear down your current session to try a new backend, which both costs time and ruins the comparison. Fazm is a native macOS app that wraps Claude Code and Codex via the Agent Client Protocol, and it is built around exactly this loop.

The anchor detail worth verifying: fazm exposes a per-chat backend switch, Claude Code through @agentclientprotocol/claude-agent-acp or Codex through codex-acp, plus a custom API endpoint field that accepts any Anthropic-compatible gateway (a corporate proxy, GitHub Copilot, or your own). So a model you spotted today can go straight into a forked window with your full prior context, while the original thread stays untouched and uncompacted.

# a new model lands. you want to try it without nuking your current thread.
$ fork this chat  ->  new window, full prior context, original untouched
$ switch backend  ->  Claude Code (ACP)  |  Codex (codex-acp)
$ custom endpoint ->  any Anthropic-compatible gateway (your proxy, Copilot)
same agent loop. no auto-compacting. session survives restart.
# run the same task in both windows. compare. keep the winner.

Because there is no auto-compacting, the full history of the evaluation stays live for the lifetime of the window, so the new backend is measured against the same decisions and constraints as the old one. And because sessions survive a Mac restart, an evaluation that spans a day or a reboot does not evaporate.

A short routine for the daily firehose

1. Skim the three live feeds. Note at most one or two drops that plausibly touch your actual stack, and ignore the rest.
2. Fork your current working chat so the candidate inherits real context instead of a toy prompt.
3. Point the fork at the new backend or endpoint and re-run a task you already finished, so you have a known-good baseline.
4. Compare on the things you ship: correctness, how far the agent got without you, and whether it could reach the browser or native app the task needed.
5. Keep the winner as your default for that kind of work. Discard the rest without guilt; there will be more tomorrow.

That loop turns the endless release stream from a source of anxiety into a cheap, repeatable experiment. The model du jour stops being a thing you have to chase and becomes one more candidate you can swap in for the cost of a click.

The counterargument: sometimes the model really is the story

To be fair, there are days when a release genuinely shifts what is possible: a meaningfully cheaper open-weight model that fits on your machine, a context-length jump that changes how much you can hand the agent at once, or a new modality. When that happens, the harness point still holds; it just cuts the other way. A real capability jump is worth the most when your loop can adopt it instantly and keep all your existing work. The harness is what lets you capitalize on a genuine breakthrough the same afternoon it ships, rather than weeks later after you have rebuilt your setup around it.