Two days in May. One open-source Mac agent. Six releases and a third backend.

Neither Hugging Face nor GitHub publishes a release list keyed to a calendar day. Both order discovery by a rolling trending score that has no notion of dates. The three live feeds worth opening on any given day are huggingface.co/models sorted by trending, huggingface.co/papers, and github.com/trending. For one open-source AI project whose own commit log carries the exact dates 2026-05-20 and 2026-05-21, the macOS agent Fazm shipped six production releases in those 48 hours: v2.9.26, v2.9.30, v2.9.31 on May 20, and v2.9.32, v2.9.33, v2.9.34 on May 21. The headline change was 18 commits on May 21 that added Gemini as a third swappable backend through the Agent Client Protocol, alongside the existing Claude Code and Codex backends. May 20 also landed cross-device phone-to-popout routing on chat.fazm.ai, a Cloudflare R2 mirror for Sparkle updates in China, and a SIGHUP-based graceful bridge restart.

Fazm already wrapped Claude Code (through claude-agent-acp) and OpenAI Codex (through codex-acp) as two interchangeable backends per chat. On May 21, 2026 the acp-bridge package gained a third: @google/gemini-cli, wrapped by a new GeminiProvider, with a probe handler in ACPBridge.swift, a .geminiProbeResult case added to the bridge protocol, GEMINI_API_KEY forwarding from the desktop into the bridge environment, and Gemini Flash plus Gemini Pro entries in the AI picker with a curated model whitelist. Every chat window can now choose Claude Code, Codex, or Gemini as its agent loop, and the choice is persisted per window through the same per-popout session machinery that landed on May 20.

Open https://github.com/mediar-ai/fazm and read three dated artifacts. CHANGELOG.json at the root lists every release with its date; the entries for 2.9.26, 2.9.30, 2.9.31, 2.9.32, 2.9.33, and 2.9.34 all carry 2026-05-20 or 2026-05-21. git log filtered to those two dates shows the commits in chronological order. The Gemini integration touches acp-bridge/package.json (the @google/gemini-cli dependency), a new GeminiProvider file, and ACPBridge.swift (the .geminiProbeResult case). The per-popout phone routing touches web/components/SessionSidebar.tsx (204 lines) and Desktop/Sources/Chat/WebRelay.swift. All commits are public.

Two reasons. First, Fazm is a wrapper around the Agent Client Protocol; every existing backend (Claude Code through claude-agent-acp, Codex through codex-acp) speaks the same protocol, so a chat window does not care which agent loop is running underneath. Wrapping gemini-cli through ACP keeps the swap mechanic uniform; you change one identifier in the bridge probe and the same UI, the same session persistence, and the same one-click forking apply. Second, gemini-cli already has its own agent loop, tool calling, workspace trust model, and file system semantics. Hitting Gemini through the Google API directly would mean rebuilding that loop inside Fazm, with two copies of every fix. The ACP route gets the maintained loop for free and adds one Mac-native UI layer on top.

Six entries in CHANGELOG.json. v2.9.26 (May 20) fixed pop-out chat windows collapsing to the same workspace after an app restart and restored the copy button on the active question bubble. v2.9.30 (May 20) added a session sidebar, model picker, workspace selector, and stream-target switcher to chat.fazm.ai for phone-to-popout routing, and added a programmatic browser-mode toggle. v2.9.31 (May 20) let the personal account chooser reuse existing Claude or ChatGPT sessions and fixed Python-based MCP loading on universal-DMG installs. v2.9.32 (May 21) re-enabled the per-tool watchdog with per-tool ceilings (Task subagent 30 min, Bash 15 min, MCP 5 min) and removed the 1M-context picker entries that were producing entitlement errors. v2.9.33 (May 21) bundled the Assrt QA testing MCP and fixed a crash when popping out chat windows during streaming. v2.9.34 (May 21) fixed selectable text and markdown going blank inside detached pop-outs.

Not a foundation model release with new public weights. The trending lists on huggingface.co/models and huggingface.co/papers carry their own daily rotation, with no notion of a calendar date built into the ranking. What did land on these two dates, in a real shipping open-source AI product, was the Gemini backend integration on May 21 and the cross-device phone-to-popout routing on May 20. The two are connected: each lets one agent loop reach further out of its sandbox, one across providers, one across devices.

Fazm runs as a native Mac app in which each chat can detach into its own pop-out window with its own working directory, agent backend, and model. After v2.9.30 on May 20, 2026, the phone web chat at chat.fazm.ai exposes a sidebar that lists every pop-out currently open on your Mac, identified by its workspace path and active model. You pick one, type from your phone, and the message routes to that specific window. Internally, a phone message now carries a sessionKey field equal to 'main' or 'detached-<uuid>', and the desktop WebRelay dispatches to that exact session. Combined with the May 21 Gemini integration, you can now send a phone message into a specific Mac window running a specific backend (Claude Code, Codex, or Gemini).

Three feeds plus one habit. huggingface.co/models sorted by trending for weights and quantized variants. huggingface.co/papers for research where an implementation is linked. github.com/trending for the application layer of agents, MCP servers, and inference engines. The habit is to read the CHANGELOG of one open-source project you actually use. The trending feeds tell you what is loud this week. A real changelog tells you what shipping actually looked like that day, including the small decisions a trending feed will never index.

No. Fazm is a native macOS wrapper that runs Claude Code, Codex, and now Gemini as agent-loop backends through the Agent Client Protocol. You bring your own Claude Pro or Claude Max, your own ChatGPT account, or your own GEMINI_API_KEY, and usage hits whichever plan you logged into. Voice input is on-device through WhisperKit. The source is at github.com/mediar-ai/fazm, and the app is free to start and runs on macOS 14 and newer.

The companion page focused on May 20 alone lives at https://fazm.ai/t/new-ai-model-releases-papers-open-source-may-20-2026. It narrows in on the 129 commits that produced v2.9.26, v2.9.30, and v2.9.31, and explains the per-popout phone routing in more depth. This page covers the next 24 hours on top of that and centers on the Gemini integration that landed May 21.