AI NewsSeptember 21, 2026

Meta opened Muse to third-party developers, letting outside services build connectors so users can reach them just by asking their agent, ten days after Muse's public launch.
Mark Zuckerberg announced on September 18 that developers can now build connectors for Muse, letting users access outside services just by asking their agent. Submissions go through Meta's review for functional, security, and legal requirements before appearing in Muse's directory.
Meta also added Stripe Link for payments, pointing toward connectors that complete transactions, not just read and summarize. It's a fast pivot: Muse launched just ten days earlier, and Meta is already racing to make it the default way people reach other apps.
Key Capabilities:

• Developer submission portal: Businesses submit connectors through a three-step review process.
• Payment-ready: Stripe Link integration points toward transactions, not just read-only tasks.
• Fast rollout: Opened to developers just ten days after Muse's public consumer launch.

Z.ai's ZCode was found silently uploading encrypted snapshots of developers' full Git histories to Alibaba Cloud with no working opt-out; the company apologized, patched the client, and open-sourced the code.
A researcher found in mid-September that Z.ai's ZCode coding tool was silently packaging users' full Git histories into encrypted archives, uploading them to an Alibaba Cloud bucket with keys held only by Z.ai. The feature had no working opt-out and wasn't disclosed anywhere.
Z.ai fixed it in version 3.14.0, removed the feature, and cleared the bucket, saying an audit confirmed no data was used for training. It apologized and open-sourced ZCode entirely, though the researcher who found the bug says deletion can't be independently verified.
Key Capabilities:

• Silent upload: Full Git histories and project files packaged and sent without disclosed opt-out.
• Server-held keys: Only Z.ai held the decryption key, so users couldn't verify what was uploaded.
• Now open-source: Z.ai released ZCode's full codebase for public scrutiny after the backlash.

An engineer paired a fast decision-making model with GPT-6 Astra's planning to defeat Minecraft's Ender Dragon in under 9 minutes for less than a dollar, open-sourcing the code behind the run.
Engineer Ronak Malde paired Jev, a fast decision-making model, with GPT-6 Astra for planning, beating Minecraft's Ender Dragon in 8 minutes 43 seconds from an empty inventory. The run cost under a dollar total: about a penny for Jev, 96 cents for Astra.
The split mirrors "thinking fast and slow": Astra sets high-level goals, while Jev makes near-instant tactical calls during combat and movement. A rival team using a different fast model, Jesse, shaved the run to 7 minutes 6 seconds. The code is open-sourced.
Key Capabilities:

• Split roles: Astra handles high-level planning; Jev makes near-instant tactical decisions.
• Dirt cheap: Total model cost under a dollar for the full run.
• Open-sourced: Full harness code is public for others to replicate or improve on.

Alibaba released Qwen3.8-LiveTranslate, a real-time interpretation model that cuts translation lag from 2.8 to 2.3 seconds across 60 languages, closing in on the pace of professional human interpreters.
Alibaba released Qwen3.8-LiveTranslate on September 19, its next-generation real-time interpretation model. A new Interleave architecture weaves audio and text into one stream, cutting average lag from 2.8 to 2.3 seconds across 60 languages, near the range human interpreters work at.
New features include real-time speaker diarization with stable voice cloning per speaker, synchronized bilingual on-screen captions, and long-context disambiguation that keeps names and terms consistent across a conversation. It's live now on Alibaba Cloud.
Key Capabilities:

• Near-human pace: Average lag drops to 2.3 seconds, in the same range as professional interpreters.
• Speaker-aware: Real-time diarization separates speakers and clones each voice for the translation.
• Live now: Available as a hosted API on Alibaba Cloud Model Studio and QwenCloud.

Alibaba released Qwen-Image-2.1, a compact 7B open-weight model for image generation and editing that natively handles transparency and up to 10 reference images, under a non-commercial research license.
Alibaba released Qwen-Image-2.1 on September 20, a 7B open-weight model that unifies image generation and editing in one package. It natively generates and edits transparent RGBA images, accepts up to 10 reference images, and runs on consumer GPUs like an RTX 3090.
It scores 60.28 on Qwen's own benchmark, beating larger rivals like Nano Banana 2.0. But there's a catch: unlike earlier Qwen-Image releases under Apache 2.0, this one ships under a research-only license that requires a separate agreement for commercial use.
Key Capabilities:

• Native transparency: Generates and edits real RGBA alpha channels, not simulated cutouts.
• Multi-reference editing: Combines up to 10 reference images for group portraits or virtual try-ons.
• Licensing catch: Open weights, but research-only; commercial use needs a separate license from Alibaba.