New AI projects on Hugging Face or GitHub, May 15 2026. Finding them is easy. Telling which ones survive is the real skill.

The question hides a wrong assumption

Most write-ups for a dated query like this one hand you a list. Five repositories, four models, a paper or two, each with a one-line description and a star count. The list looks authoritative because it is specific. It is also close to useless a week later, and it was never the thing you actually needed.

What you needed was a way to decide. Out of everything that appeared around May 15, 2026, which projects are worth installing, depending on, or building against, and which are demos that already peaked? A static list cannot answer that. Trending is a popularity measurement taken over a rolling window. A coordinated launch, a single viral post, or a well-timed announcement spikes it cheaply, and the spike decays within days. The list you found is a photograph of a moment that has already passed.

The signal that does survive is cadence. Not how much attention a project earned on one day, but how consistently it ships over weeks. Cadence is hard to fake, it is recorded in artifacts the maintainer authors directly, and it correlates with the only thing you actually care about: will this project still exist when you need it.

Two ways to read a new project

These are not opposites you choose between. A trending feed is how you discover a project; cadence is how you decide whether to trust it. The mistake is stopping at the first one. Here is what each actually measures.

Five checks before you run anything you found this week

None of these require installing the project. They are all things you can read in a browser tab in about two minutes. The first four are the qualifying checks. The last two are the most common patterns that should make you close the tab.

Worked example: what May 15, 2026 looked like inside one project

Fazm is a native macOS agent, open source under the MIT license, that wraps Claude Code and Codex in a desktop UI. It is a useful specimen for this exercise for one boring reason: its release history is a plain JSON file at the root of the repository, so the cadence check takes seconds. Here is the slice of that file carrying the date in this question.

Four releases in one day. Look at what they are. A freeze when opening a tab. A CPU spike from an animation running in a hidden window. Stale session data left behind by a closed window. A misleading error message from the Codex backend. False billing errors triggered by a Claude 529 overload response during an Anthropic outage. Not one of these is a feature. Every one is a specific, observed, named defect, fixed and shipped the same week it was found.

That is exactly the texture you are looking for. Headline features are easy to fake with a launch post. A steady stream of small, dated, unglamorous fixes is not, because it can only come from a project that has real users hitting real edges. The week of May 11 to 17, 2026 shows three separate release days in this one file, with four releases on the 15th alone.

The second file worth opening: the dependency manifest

After the changelog, the highest-signal file in any AI project is the dependency manifest. It tells you what the project is built on and how carefully the author tracks it. In Fazm the relevant one is acp-bridge/package.json, the manifest for the Node sub-process that connects the macOS app to the agent runtime. As of mid-May 2026 it pins @playwright/mcp at 0.0.73, @agentclientprotocol/claude-agent-acp at 0.29.2, @zed-industries/codex-acp at 0.12.0, and the usual small libraries (ws, zod).

Exact version pins on fast-moving agent infrastructure mean the author updates deliberately and tests against a known set. The opposite pattern, wildcard ranges or vendored copies of upstream packages frozen a year ago, tells you a project has stopped tracking the ecosystem it depends on. For a new project you are evaluating, the manifest answers a question the README never will: is this still being steered.

Why most "new AI projects" are statistically noise

The instinct to find more new projects is the wrong instinct. The catalog is already enormous and brutally concentrated. Hugging Face's Spring 2026 State of Open Source report put hard numbers on it.

Roughly half of all models on Hugging Face have fewer than 200 total downloads. The top 200 models, about 0.01 percent of the catalog, take 49.6 percent of all downloads. The distribution is not a gentle curve, it is a cliff. Most of what appears on a trending feed on any given day is a real upload that will never be adopted and never maintained past its first week.

This is why a dated roundup is the wrong tool. It treats every entry as equal because they all appeared in the same window. The skill worth building is the opposite of collecting: it is discarding fast. A two-minute cadence and manifest check throws out the noise before you waste an afternoon on it.

When the trending list is genuinely enough

The cadence check is not always worth running. If you are browsing for awareness, skimming what the field is excited about this week with no intention of depending on any of it, a trending feed is exactly the right amount of effort. The same is true for projects from organizations with a long, public track record: a new release from an established model lab does not need its commit log audited before you try it.

The cadence and manifest checks earn their two minutes when the decision has a cost: you are about to add a dependency, wire a tool into a workflow, or recommend something to other people. That is the moment the question stops being "what is new" and becomes "what will still be here." The two halves of this guide map cleanly onto those two questions. The companion piece on Hugging Face or GitHub for new AI projects in April 2026 covers the discovery half, the split between weights and code, in more depth.