Latest AI model releases, papers, and open-source projects (May 21 to 22, 2026)

There is no canonical dated index of AI releases for a specific 48-hour window in May 2026, on any platform. The honest answer is three rolling feeds plus one project changelog you actually trust. Here is what that looks like for May 21 to 22, with a verified worked example.

Matthew Diakonov, Written with AI

Published May 22, 20269 min read

Direct answer (verified 2026-05-24)

No platform publishes a 48-hour AI release index for May 21 to 22, 2026. The verifiable answer is to combine three rolling feeds with one or two project changelogs you actually run. As a worked example: the open-source macOS agent Fazm shipped five releases in those 48 hours, recorded with ISO dates in CHANGELOG.json at github.com/mediar-ai/fazm: v2.9.32 and v2.9.33 (May 21) added an opt-in Assrt QA testing MCP and fixed per-session tool tracking; v2.9.34 (May 21) fixed pop-out window rendering; v2.9.35 and v2.9.36 (May 22) added Google Gemini Flash and Gemini Pro as selectable backends alongside Claude and ChatGPT, with Gemini wired up as a free option when built-in credits run out.

The three feeds that actually carry the signal

Every aggregator that promises a dated 24-hour or 48-hour roundup of AI releases is downstream of the same three sources. Going to those sources directly is faster and more honest than waiting for a newsletter.

huggingface.co/papers/trending ranks new research papers by recent attention. Most entries link to code. This is the right feed for new method papers, fine-tuning techniques, and benchmark results.
huggingface.co/models?sort=trending ranks model weights by popularity over the last week. This is where you see new base models, fine-tunes, quantized GGUF variants, and adapter releases.
github.com/trending ranks repositories by recent star velocity. This is where the agent harnesses, MCP servers, inference engines, and tooling projects surface.

None of these are calendar-keyed. They are rolling popularity scores. To check what actually shipped in a specific 48-hour window, you have to open a project and read its dated changelog or Releases page.

Method vs anti-method for a dated 48-hour AI release window

The dated-index expectation collapses on first contact with reality. Use the methods that actually exist.

Feature	Method that does not	Method that works in May 2026
Frontier model drops (OpenAI, Anthropic, Google)	Hourly trending refresh	Lab blog + weekly skim
Research papers with linked code	Dated 24h paper index (does not exist)	huggingface.co/papers/trending (daily)
Open weights + quantized variants	Calendar-keyed model release feed	huggingface.co/models?sort=trending
Harnesses, MCP servers, runtimes	Single aggregated 24h GitHub index	github.com/trending + per-project CHANGELOGs
Agent-layer fixes for the past 48h	Wait for a roundup newsletter	Read a maintained project's dated CHANGELOG.json

The Fazm changelog for May 21 to 22, 2026: five releases, one structural change

Fazm is an open-source native macOS agent that wraps Claude Code and Codex via ACP. Its CHANGELOG.json records every release with an ISO date. Reading the five entries dated 2026-05-21 and 2026-05-22 gives a concrete, verifiable picture of where AI agents were actually shipping in that 48-hour window.

Five Fazm releases across May 21 to 22, 2026

v2.9.32 (May 21)

Added Assrt QA testing MCP (opt-in). Fixed per-session tool tracking so one slow MCP call no longer freezes sibling chats. Re-enabled the tool watchdog with safer per-tool ceilings (Task 30m, Bash 15m, MCP 5m). Stopped dropping post-close notifications. Removed the 1M-context Sonnet/Opus picker entries that could trip an entitlement error.

v2.9.33 (May 21)

Hardened Assrt: imports the user's real browser session (cookies, localStorage, IndexedDB) into managed browsers when enabled. Made error messages in Browser Automation cards selectable. Fixed an empty chat canvas during streaming. Fixed a crash when popping out or closing windows mid-stream. Fixed Settings > Import sessions failing with 'ai-browser-profile not bundled' on universal builds.

v2.9.34 (May 21)

Fixed selectable chat text going blank or crashing in detached pop-out windows. Fixed markdown content rendering blank in detached pop-out windows.

v2.9.35 (May 22)

Added Google Gemini Flash (latest) and Gemini Pro (latest) as selectable models in the AI picker alongside Claude and ChatGPT. Fixed Gemini Flash and Gemini Pro not being able to call MCP tools (screen capture, browser automation, WhatsApp). Clarified extension-connection-test guidance to keep a regular webpage open in the active tab. Aligned Gemini labels with the tier-first naming convention.

v2.9.36 (May 22)

Added Google Gemini as a free option when the built-in AI credits run out, alongside Connect Claude and Connect ChatGPT. This is the structural change of the 48-hour window: three model families, three bring-your-own-account paths, all driving the same agent loop with the same MCP tool surface.

The anchor fact: v2.9.36 made Gemini a first-class backend

The release that defined the 48-hour window was not the largest, by line count. v2.9.36 has a single changes entry. But it is the entry that closed the capability gap that had been open since v2.9.0. Before May 22, the Fazm agent loop ran on Claude Code via ACP and Codex via ACP, both bring-your-own-account, with a built-in Claude credit pool as the free fallback. When that pool ran out, the user had to connect their own paid account or stop.

v2.9.36 added Google Gemini as a third free option, with the same MCP tool surface (screen capture, browser automation, WhatsApp, accessibility-driven Mac control). v2.9.35 earlier the same day fixed Gemini Flash and Gemini Pro not being able to call MCP tools at all, which is the kind of integration-layer bug that distinguishes "model is announced" from "model is usable inside an agent loop".

The exact entries are in CHANGELOG.json. You can verify them yourself:

CHANGELOG.json (truncated to May 21 to 22, 2026)

Source: github.com/mediar-ai/fazm, accessed 2026-05-24.

The agent-layer failure categories visible in this 48-hour window

The Fazm releases on May 21 to 22 are not Fazm-specific. They are patches for failure categories that any agent built on top of Claude Code, Codex, or now Gemini eventually has to handle. Five of them land in this window.

Cross-session tool-call deadlock. v2.9.32 fixed pop-out chat windows getting stuck on tool calls when another session was busy. Tool activity is now tracked per-session, so one slow tool no longer freezes every other open conversation.
Watchdog runaway. v2.9.32 re-enabled the tool watchdog with safer per-tool ceilings: Task subagent 30 min, Bash 15 min, MCP 5 min, default 10 min. Genuinely hung tools now auto-recover instead of leaving the chat spinning forever.
Dropped post-close notifications. v2.9.32 stopped silently dropping config, mode, and usage updates that arrived after a chat session closed. Previously these could leave the UI showing stale model or mode info.
Wrong-entitlement model picker entries. v2.9.32 removed the 1M-context Sonnet and Opus picker entries that could land users on a variant their Claude account did not have access to, surfacing a hard usage-credits error with no recovery.
MCP tool calling on non-Anthropic models. v2.9.35 fixed Gemini Flash and Gemini Pro not being able to call MCP tools at all. Until you patch this, "Gemini in your agent" is just a chat model with no hands.

Every other wrapper, IDE plug-in, and computer-use agent shipped on top of the major model APIs in May 2026 has to face the same five. That is the value of reading a single project changelog over 48 hours instead of a roundup: you see where the ecosystem is actually breaking, not where the press releases are aimed.

What about papers from May 21 to 22, 2026?

Hugging Face does not publish a calendar-keyed paper index for either date. The closest thing is the rolling list at huggingface.co/papers/trending, which surfaces papers by recent attention rather than upload date. If you want a 48-hour cut, the working method is to open the trending list and filter the entries whose first-version arXiv date lands inside your window. Most of what you will see on May 21 to 22 is not net-new method papers but evaluation notes, follow-up work on the previous week's frontier-model drops, and benchmark updates.

What about open-source projects from May 21 to 22, 2026?

The right move is to bookmark github.com/trending and github.com/trending?since=daily, then open the README and the Releases page of any project that looks like it could land in your stack. The signal is in the dated tags, not the star count. The Fazm tag history at github.com/mediar-ai/fazm/releases is one example of what that record looks like in practice for May 21 to 22, 2026.

A practical 48-hour AI release routine

The routine that holds up in May 2026, applied to the May 21 to 22 window:

Skim the three rolling feeds (papers/trending, models/trending, github/trending). Five minutes.
Read the lab blogs you care about. None of OpenAI, Anthropic, or Google DeepMind made a frontier-model announcement on either of these two dates, so this step is short.
Open the CHANGELOG of one or two projects you actually run. This is where the application-layer signal lives. For May 21 to 22, the Fazm CHANGELOG is the worked example.
Write down one ecosystem-level pattern you noticed. For this window, the pattern is: the agent layer is normalizing on multi-backend (Claude, ChatGPT, Gemini), and the integration bugs are the cost of admission.

Want a 15-minute walkthrough of the multi-backend agent loop?

If you are building or evaluating an AI agent setup for your own machine, book a call. We will open Fazm together and show you the Claude, ChatGPT, and Gemini path on the same MCP tool surface.

Frequently asked questions

What AI model releases, papers, and open-source projects shipped specifically on May 21 to 22, 2026?

No platform publishes a dated 48-hour index for either day, so there is no single canonical answer to copy. The honest method is the same: read huggingface.co/papers/trending, huggingface.co/models?sort=trending, and github.com/trending, then verify any specific project against its own dated changelog. A verifiable data point for exactly those two dates: the open-source macOS agent Fazm shipped five releases stamped 2026-05-21 and 2026-05-22 in its CHANGELOG.json. v2.9.32 and v2.9.33 on May 21 added an opt-in Assrt QA testing MCP and fixed per-session tool-watchdog behavior. v2.9.34 on May 21 fixed pop-out chat rendering. v2.9.35 and v2.9.36 on May 22 added Google Gemini Flash and Gemini Pro as selectable backends alongside Claude and ChatGPT, with Gemini wired up as a free option when built-in AI credits run out.

Where do I actually find AI model releases, papers, and open-source projects in a 48-hour window like May 21 to 22, 2026?

Three rolling feeds, none of which are calendar-keyed. huggingface.co/papers/trending lists new research, usually with linked code. huggingface.co/models?sort=trending shows weights, quantized variants, and fine-tunes by popularity in the last week. github.com/trending surfaces harnesses, MCP servers, agent runtimes, and inference engines. To check whether any specific project is actually shipping in that window, open its CHANGELOG or its Releases page on GitHub and read the dated entries directly. The Fazm example for May 21 to 22 is fully verifiable: every release is in CHANGELOG.json at the root of github.com/mediar-ai/fazm with an ISO date field.

Did anything notable happen at the major AI labs on May 21 or May 22, 2026?

Neither OpenAI, Anthropic, nor Google DeepMind made a major frontier-model announcement on either of those exact dates. The May 21 to 22 activity worth recording is at the agent and tooling layer, where the gap between a published model and a model you can actually use through an agent loop is still measured in days. Fazm's v2.9.36 release on May 22 is a concrete example: Google Gemini Flash and Gemini Pro had been available via API for some time, but May 22 is the day they became a one-click backend inside a native Mac agent with full MCP tool access, alongside Claude (Code) and ChatGPT (Codex). Every entry is dated in CHANGELOG.json at github.com/mediar-ai/fazm.

Why was the v2.9.36 Gemini backend release significant for the May 21 to 22 window?

Because it closed a real capability gap on a real cadence. Before v2.9.36, the Fazm agent ran on Claude Code via ACP and Codex via ACP, both bring-your-own-account, with a built-in Claude-credit pool as the free fallback. When that pool ran out, the user had to connect their own paid account or stop. v2.9.36 added Gemini as a third free option, with the same MCP tool surface (screen capture, browser automation, WhatsApp, accessibility-driven Mac control). The earlier v2.9.35 release on the same day also fixed Gemini Flash and Gemini Pro not being able to call MCP tools at all, which is the kind of integration-layer bug that distinguishes 'model is announced' from 'model is usable in an agent'.

What does the Assrt QA testing MCP in v2.9.32 and v2.9.33 do?

Assrt is an AI-powered QA testing MCP that runs real browser tests against web applications. v2.9.32 added it as an opt-in tool in Settings > Browser Automation, bundling test scenarios alongside the existing browser tools. v2.9.33 hardened the integration: it now imports the user's real browser session (cookies, localStorage, IndexedDB) into both managed browsers when enabled, and surface error messages in the Browser Automation cards as selectable text so they can be copied for support. This is the agent-tooling pattern that defined the May 21 window: not new models, but new MCP servers that let the existing models reach further into the user's actual machine state.

What recurring AI-agent failure categories did the May 21 to 22 releases address?

Five, all visible in CHANGELOG.json. First, cross-session tool-call deadlock: v2.9.32 fixed pop-out chat windows getting stuck on tool calls when another session was busy; tool activity is now tracked per-session, so one slow tool no longer freezes every other open conversation. Second, watchdog runaway: v2.9.32 re-enabled the tool watchdog with safer per-tool ceilings (Task subagent 30 min, Bash 15 min, MCP 5 min, default 10 min) so genuinely hung tools auto-recover. Third, dropped post-close notifications: v2.9.32 stopped silently dropping config, mode, and usage updates that arrived after a chat session closed. Fourth, wrong-entitlement model picker entries: v2.9.32 removed the 1M-context Sonnet and Opus picker entries that could land users on a variant their Claude account did not have access to. Fifth, multi-tool MCP integration on non-Anthropic models: v2.9.35 fixed Gemini Flash and Gemini Pro not being able to call MCP tools at all.

Where can I check each of those changes against a primary source?

Two sources. The on-disk source of truth is CHANGELOG.json in any local clone of the public Fazm repo. The public mirror is https://github.com/mediar-ai/fazm/blob/main/CHANGELOG.json, which renders the same content. Each release object has a version field, an ISO date field, and a changes array. Versions 2.9.32 through 2.9.36 with dates 2026-05-21 and 2026-05-22 are present at the time of writing. The release feed at https://fazm.ai/download serves the matching notarized macOS builds. Nothing on this page asks you to take the changelog on faith.

What is the right cadence to check for new AI releases if 24 hours is the wrong unit?

It depends on what you are trying to track. For frontier model releases (OpenAI, Anthropic, Google DeepMind, Mistral, Meta, the major Chinese labs), once a week is more than enough; major drops are announced on the lab's own blog and propagate within hours. For Hugging Face papers and code, daily skimming of huggingface.co/papers/trending catches what matters; sub-daily polling is mostly noise. For the agent and tooling layer (Claude Code, Codex, Cursor, Continue, MCP servers, computer-use harnesses), watching a small number of project changelogs on a rolling 7-day window beats any trending feed. The Fazm worked example for May 21 to 22 is that a single project's dated record, read over 48 hours, often reveals more about ecosystem maturity than any aggregator does.

Does Fazm itself help me answer 'what shipped in the past 48 hours' from inside the app?

Yes, through the deep-research skill that auto-installs to ~/.claude/skills/deep-research/SKILL.md on first launch. The skill runs an 8-phase pipeline (Scope, Plan, Retrieve, Triangulate, Synthesize, Critique, Refine, Package), launches parallel web searches plus parallel research subagents, verifies citations where possible, and writes a markdown plus HTML plus PDF report into a dated folder under ~/Documents. Because it runs as the local Claude Code (or, after v2.9.36, Gemini or Codex) agent on your machine, the answer is grounded against fresh searches from your IP, not a cached newsletter summary.