unsloth v0.1.911-beta — Scheduled + MiniMax-M3.1-Flash + Perf Gains

Unsloth, an open-source app for running and fine-tuning AI models on your own hardware, adds scheduled tasks in this release. A user picks a prompt and a schedule (once, daily, weekdays, weekly or monthly), and each run's answer is saved as a new chat, according to the release notes. A task can use the model already loaded, another model on the machine, or one from a connected API provider. Two limits shape what it is good for: tasks run only while Unsloth is open, and they run without tools or web search, so a scheduled prompt can only work from what the model already knows and what the prompt itself contains.

The release also fixes a sandbox approval gap. Unsloth can let a model run Python or terminal commands, and a call that names a file path outside the sandbox's allowed folders waits for a person to approve it before it runs outside the operating system's sandbox. The pull request explains that the approval endpoint accepted an approval from any authenticated caller, including an app using an Unsloth API key, so such a caller could approve its own waiting call and get the unsandboxed launch that other checks refuse. Approvals from API-key callers are now rejected unless the server operator explicitly turns on full API access, while denials are still accepted from anyone. The PR cites no CVE, and its author tested on Linux with the operating system's sandbox stubbed out. Anyone who gives other apps an Unsloth API key has reason to update.