Hugging Face or GitHub for new AI projects on April 29, 2026: three things shipped, and the most useful one was not a model.
April 29 produced two model artifacts everyone covered, and one consumer-agent activity nobody covered. IBM Granite 4.1 (3B, 8B, 30B, Apache 2.0) landed on Hugging Face. QwenPaw v1.1.5 shipped on GitHub. And in the same 24 hours, Fazm pushed four consecutive patches (v2.6.3, 2.6.4, 2.6.5, 2.6.6) to its consumer macOS agent, each one a reliability fix for the agent loop. The model drops are background. The patch cadence is the story.
What launched on April 29, 2026?
- Hugging Face: IBM Granite 4.1 (3B, 8B, 30B variants, Apache 2.0, dense decoder, updated post-training pipeline for tool calling and instruction following). Confirm at huggingface.co/ibm-granite/granite-4.1-8b.
- GitHub: Alibaba's QwenPaw v1.1.5 (formerly CoPaw, rebranded April 12). Confirm at github.com/agentscope-ai/QwenPaw.
- GitHub, consumer side: Fazm shipped four consecutive releases the same day (v2.6.3 to v2.6.6), all reliability fixes for the macOS agent loop. Confirm at github.com/mediar-ai/fazm.
The headline coverage misses the operational story
Every roundup written about April 29, 2026 starts with the model drops. They are easy to write about. Apache 2.0 license, parameter count, context window, benchmark column on the model card. Granite 4.1 in particular is a credible release: the 8B variant is a usable Apple Silicon model in 4-bit GGUF on a 16 GB Mac, and the updated post-training pipeline noticeably improves tool calling for agent workflows. QwenPaw v1.1.5 is a smaller signal but confirms that Alibaba is investing in the desktop assistant surface around the Qwen model family.
What the roundups miss is that none of those weights run an agent on their own. An agent is a loop: prompt, model call, tool result, prompt, model call, tool result, repeat. The loop runs on a machine the user owns, against a UI the user is watching, talking to apps the user logged in to. Every interesting failure mode lives in that loop, not in the model. And on April 29, the loop got four discrete patches in one day on a consumer Mac agent shipping to production users. That is the part of the day worth opening.
The four patches, in the order they shipped
Each entry below maps to a real release tag and a real CHANGELOG line. The wording in the "detail" field is paraphrased, but the underlying fix is one line from the JSON.
Pop-out chat windows freezing and a crash on launch
Each pop-out used to re-render on every streaming token from every other open session. With four pop-outs running, the cross-chatter could pin the main thread, freeze the copy-on-hover affordance, and occasionally crash when a new pop-out was spawned mid-stream. The fix scopes streaming updates to the session that owns them.
Cross-chat response bleed and an OAuth reconnect flood
An AI response in one chat could leak into a following chat after a timeout, returning a duplicate of the previous turn. Separately, restarting OAuth would trigger a reconnect storm that prompted the user to re-authenticate Claude repeatedly. Both fixed.
Streaming follow-ups paired with the wrong question
If you sent a follow-up while the previous response was still streaming, the pop-out could pair the new AI message with the previous user message and orphan the new question. The fix gates assignment on stream completion. Also: a visible App Management permission warning in Settings, with a dismiss button.
Agent loops stuck on tool-call timeout
The agent could hang forever when a tool call timed out, because the synthetic timeout response was queued behind a session that had already moved on. The session now auto-interrupts on tool timeout and skips the synthetic response if the tool actually completed. The New Chat Window button also got a contextual count.
The anchor fact, in its native format
CHANGELOG.json is the source of truth. The file lives at the root of the repo and is updated by the release script in the same commit that signs the build for distribution. Below is the exact slice for April 29, 2026, four entries in a row, sorted newest first.
Verify the full file at github.com/mediar-ai/fazm/blob/main/CHANGELOG.json. Search the file for the literal string "2026-04-29" and you land on the four entries above.
How to read April 29 if you actually run an agent
There are two ways to consume a day in AI. The headline framing is optimized for shareability. The operator framing is optimized for what you would do tomorrow morning.
Two readings of the same 24 hours
IBM and Alibaba shipped on April 29. Granite 4.1 is a credible Apache 2.0 family; QwenPaw is a sign of life in the open desktop agent space. Add both to a list, link to the model cards, move on.
- lists model weights
- links benchmark numbers from the model card
- stops at the upstream artifact
The honest case for caring about the model drops anyway
The argument above sounds like it dismisses the model side. It does not. Granite 4.1 8B is a real improvement on Granite 4.0 for agent workflows, specifically because the post-training pipeline was tuned for cleaner tool-use JSON. On a Mac running the agent loop locally, that translates to fewer rejected tool blocks and fewer retries, which means the harness has fewer failure modes to recover from. The harness improvements on April 29 do not replace the model improvements; they compound with them.
QwenPaw is a similar story on a smaller scale. The April 12 rebrand was a positioning move toward a small-and-large model collaboration pattern: a tiny local model handling tool routing, a frontier model handling planning. That pattern matters because it lets the consumer Mac side do more of the work locally without pretending a local 8B is a planner. v1.1.5 on April 29 was a patch on top of that line, not a thesis change, but the line itself is worth tracking.
The point is not that model drops do not matter. The point is that on any given day, the marginal reader is better served knowing what shipped on the harness side too, because that is where most of the daily failure surface lives. Most pages will never tell you that.
Why the four patches all live in the accessibility path, not the vision path
Notice what was not in any of the four releases: a fix to a screenshot pipeline, a vision-model prompt change, an OCR bug, a pixel-matching threshold tweak. That absence is not random. Fazm drives the user's machine through the macOS accessibility tree, which is a structured representation of every UI element the frontmost app exposes: buttons, text fields, list rows, menus, labelled values. The agent reads the tree as data and operates on it directly. It does not screenshot a window and ask a vision model what the window says, because the accessibility tree already says it.
That choice is also why the four April 29 patches were possible. Pop-out chat window state, OAuth reconnect, paywall message persistence, tool-call timeout watchdog: every single one is a piece of state held in Swift on the local machine, not in a model prompt. Local state you can fix. Model behavior you can only constrain. The accessibility path keeps the failure modes on the local-state side of the line, which is the side where four releases in 24 hours is realistic. Vision-first agents have an entirely different daily bug surface, dominated by prompt and pixel concerns that are slower to reproduce and slower to ship a fix for.
What to do with this on April 30 and after
If you came here looking for "what shipped on April 29 2026, on Hugging Face or GitHub", the answer is three items in three different categories, given above. If you came here because you actually run an agent, here is the more useful answer.
- Pull Granite 4.1 8B in 4-bit GGUF and benchmark it locally on your own workflow, not the model card's. The tool-call pipeline is the part that improved; that is the part that matters for an agent.
- Read the April 29 CHANGELOG.json slice from a real consumer agent (Fazm's above will do, or your own). Pay attention to what the fix touches: state, not prompt. That is where you should be putting your own engineering time.
- Subscribe to harness changelogs the same way you subscribe to model release feeds. Most readers do not, which is why operator framing of any given day is rare.
- When you compare agents for your team, ask for a 30-day patch count on the consumer build, not a feature list. Four releases in a day, four times in a month, is a different product than one release a month.
Want the consumer-agent perspective for your team?
Twenty minutes on the harness side of running AI agents in production on macOS, with the exact failure modes we saw on April 29 and how we ship a fix the same day.
Frequently asked questions
What actually launched on April 29, 2026 across Hugging Face and GitHub?
Three concrete things. IBM Granite 4.1 (3B, 8B, 30B variants, Apache 2.0, dense decoder, 131K context with claims of 512K, an updated post-training pipeline for tool calling and instruction following) landed on Hugging Face under ibm-granite. On GitHub, Alibaba's agentscope-ai team shipped QwenPaw v1.1.5, the first dot release after the CoPaw to QwenPaw rebrand on April 12. And on the consumer side, Fazm pushed four releases in 24 hours (v2.6.3, 2.6.4, 2.6.5, 2.6.6), all reliability fixes for the agent loop, visible in CHANGELOG.json at the root of github.com/mediar-ai/fazm. Model weights and a consumer app shipping the same day, both worth opening, but each tells a different story.
Why does the Fazm four-in-24-hours cadence matter more than another model drop?
Because new weights are downstream. Until somebody wires them into a real harness on a real machine running a real workflow, they are a tar file. The patch sequence on April 29 was the operational tax of running an agent loop on top of frontier models, written down in plain English. Pop-out chat windows freezing when multiple sessions stream tokens at the same time. Tool-call timeouts leaving the agent stuck. OAuth re-auth prompts triggered by reconnect floods. Unsent messages disappearing behind a paywall sheet. None of that is in any model card. All of it is what the user actually sees when they put weights into production. That is the bug surface, and on April 29 you could watch it shrink in real time.
Where can I verify the four Fazm releases for April 29 myself?
Open https://github.com/mediar-ai/fazm/blob/main/CHANGELOG.json in a browser. The file is a JSON object with two keys, unreleased and releases. The releases array is sorted newest first. Search for the literal string "2026-04-29" and you will find four entries in a row: 2.6.6, 2.6.5, 2.6.4, 2.6.3. Each has a changes array listing the specific fix. The whole file is updated by the release script and signed into git at the same time the macOS app is signed and notarized, so the JSON the page references and the version that downloaded to a user's Mac are the same artifact.
Does Hugging Face publish anything that talks directly to Fazm's agent loop?
Not directly. Hugging Face publishes weights. A Mac agent talks to weights through an inference server (llama.cpp, MLX, vLLM, Ollama) and a translation layer (LiteLLM, claude-code-router, or any Anthropic-compatible adapter). Fazm's Swift code routes the agent process through whatever endpoint the user puts in the Settings field labelled customApiEndpoint. That field is the seam. Read Desktop/Sources/Chat/ACPBridge.swift in the Fazm repo, find the lines that pick up customApiEndpoint and set ANTHROPIC_BASE_URL on the spawned agent process, and that is where a Hugging Face model published April 29 becomes a model the agent talks to.
Is Granite 4.1 worth pulling on a Mac with 32 GB or less?
Granite 4.1 8B in 4-bit GGUF fits comfortably on an M-series Mac with 16 GB and runs at a usable rate for tool-call-heavy workflows. The 30B variant is closer to a 32 GB or 48 GB machine in 4-bit. The improvements over 4.0 are mostly on tool calling and instruction following, which is the part that matters for an agent rather than a chat app. The release notes describe an updated post-training pipeline aimed at exactly that. For Fazm specifically, the bottleneck is not the model size, it is whether the model produces well-formed Anthropic tool_use blocks under streaming load, and Granite 4.1 is the first Granite that does that out of the box without a router in the middle rewriting bad JSON.
What is QwenPaw and why is the v1.1.5 release noteworthy?
QwenPaw is Alibaba's open-source desktop assistant project, formerly CoPaw, hosted at github.com/agentscope-ai/QwenPaw. The April 12 rebrand was a positioning move: tighter coupling with the Qwen model family and a sharper focus on the model layer, including the small-and-large collaboration pattern where a tiny tool-router model handles the boilerplate and hands the hard parts to a frontier model. v1.1.5 on April 29 was a patch release on top of that line. Worth noting if you care about how the open Mac and Linux desktop agent space evolves around the Qwen ecosystem, less interesting if you are already on a Claude or Anthropic-API-compatible stack like Fazm runs by default.
Why ship four versions in 24 hours instead of one bigger release?
Because the cost of a small release that lands one fix is lower than the cost of a bigger release that lands four and creates a regression. Each Fazm version that day fixed one specific user-visible problem. 2.6.3 fixed pop-out re-renders and an unrelated crash on launch when popping a new chat. 2.6.4 fixed cross-chat response bleed after a timeout and an OAuth re-auth flood. 2.6.5 fixed pop-outs pairing AI responses with the wrong question during streaming and added an App Management permission warning. 2.6.6 added a context-aware New Chat Window button and an auto-interrupt on tool timeout so the agent loop unblocks. Bundling those would have meant testing every combination against every other; shipping them in a row meant a tighter blast radius if any single one regressed.
Where does the patch cadence break down for a desktop AI agent?
Two places. One, anything that has to be code-signed and notarized. Even on a hot patch day, every release goes through Apple's signing service, and an aborted notarization round-trip can stall the next push by an hour. Two, the App Store path. Apple's review queue is the slowest link in the chain by an order of magnitude, so a consumer Mac app that ships through the store hits a hard floor of one release per workweek no matter how disciplined the engineering side is. Fazm distributes a direct-download notarized build outside the store, which is the only reason four releases in 24 hours is even possible. The same product on the store would have shipped one of the four that day and held the others.
How is this page different from the broader April 2026 roundup on the same site?
The broader page enumerates everything that shipped across the month and shows how Fazm's bridge pins agent infrastructure packages by version. This page is one day. The argument is also different: across a whole month, the model drops dominate the story because each one extends the capability surface. On a single day, the model drops are background and the real story is the operational cadence of the consumer agents that depend on them. If you want the macro picture, read /t/hugging-face-or-github-new-ai-projects-april-2026. If you typed the literal date into a search box because you wanted to know what happened on April 29, this page.
Where does Fazm fit in the broader local-versus-hosted argument on April 29?
Fazm runs the agent loop locally and talks to whichever model endpoint you point it at, hosted Anthropic by default, any Anthropic-compatible local proxy when you set customApiEndpoint. The four April 29 patches all sat in the local part: pop-out window state machine, OAuth client refresh, paywall sheet state, tool-call timeout watchdog. None of them changed the model side. That is the load-bearing observation: on a frontier-model agent, almost all of the day-to-day production work is on the harness, not the model. If you spend a week reading model release notes and never read harness release notes, you are reading the wrong half of the system.
Adjacent reading
Hugging Face or GitHub for new AI projects in April 2026, the full-month view
The whole month, including how Fazm's acp-bridge pins agent infrastructure packages by exact version.
Best new Hugging Face models for Mac agents, April 2026
The four April model releases that pass the agent filter, plus the Swift seam that wires them in.
New LLM releases April 2026 for Mac desktop agents
The narrower cut of April releases sized to run locally on Apple Silicon hardware.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.