What's new

Plain-language notes, newest first. Every line here maps to real commits in the open git history — that's the unabridged version.

July 2026

• Quenderin is on the App Store — free, for iPhone. The native SwiftUI app is live worldwide (175 regions) — fully offline, on-device, our first public store release. The Mac App Store build follows once its review clears.
• v0.2.0 — stores + dual Windows channel. Version aligned across desktop, Android, and Apple. Windows: keep free GitHub .exe installs and ship Microsoft Store (AppX) for branding and no SmartScreen — owner guide in the repo. macOS: public product is the native free Mac App Store app only (not Electron). Reliability and file-chore tools landed in the open source: batch organize, multi-file collect for reports, stronger skill memory, GUI self-checks.
• The local computer-use agent grew up — a private alternative to cloud agents. On a Mac, quenderin do "<anything>" can now click and type in any app via accessibility (--gui: buttons, menus, navigation), run any of your Apple Shortcuts, and create files as well as organise them. What a cloud agent can't offer, and we now do: undo a whole task even in a new session (quenderin undo), a per-task local audit trail (quenderin history), and a true dry run that shows exactly what it would do while touching nothing (--dry-run). It reads for real but never changes anything without your say-so. quenderin capabilities lists the whole toolkit. Why local beats the cloud here.
• Reliable on a small local model. A weak on-device model can get stuck repeating itself or emit malformed output; the agent now recovers from both — nudging once, then stopping cleanly rather than looping forever — and streams each step live so you watch it work and can stop it the instant something looks wrong.
• The agent operates your computer — locally. Grant a workspace folder and the agent can list, move, rename, and trash-to-a-visible-folder — proposing a plan you approve once, with per-step undo and a plain-text ledger of everything it did and tried. Nothing overwrites, nothing deletes, nothing leaves the machine. The announcement.
• Attach documents in chat. Drop a text file or (on Mac/iOS) a PDF into the composer and ask about it — follow-ups keep it in context. Binary files are refused with a reason, never mangled.
• One safety blocklist everywhere. The actions the agent must never take autonomously are now a single canonical list, enforced identical across all platforms by CI.
• The CLI. quenderin chat — a streaming terminal REPL plus a pipe mode (git diff | quenderin chat -p "review this") that makes code review private by construction. The announcement includes the two bugs we hit shipping it.
• Windows & Linux downloads. First public desktop preview (v0.1.0): installer and portable .exe, AppImage, .deb — built in public by CI, linked on the download page.
• The consent gate. Every platform now asks for one explicit agreement before first use: AI output is your call to verify, not advice to follow.
• The roadmap went public. Where this is all going, stage by stage: quenderin.org/roadmap.
• Gemma 4 12B joins the library. Google's newest open model (Apache 2.0, 7.4 GB download) — the strongest multilingual option we ship. Recommended automatically on 16 GB machines.
• About is a page now. Clicking the elf opens a proper About screen instead of a popup menu; help, changelog, and privacy land on their quenderin.org pages.
• Daily-use fixes. Conversation switching is instant again (a long-press gesture was eating clicks — removed). Chat bubbles come in four artwork-derived colors. Sheets close when you click outside them; no more hunting for Done.
• The window chrome behaves. The traffic lights sit on one surface instead of straddling a color seam.

June–July 2026 — the credibility round

• Repetition walls died. A model repeating the same paragraph three times kills trust instantly, so it's now stopped in three places: a repetition penalty in the sampler, a mid-stream loop detector, and a final collapse pass. The full list of what can still go wrong is published.
• Streaming that keeps up. The transcript follows generation like you'd expect, stops following the moment you scroll up, and offers a ↓ jump back. A Stop button actually stops. Cyrillic and emoji arrive intact even when the model splits them across tokens.
• A real model library. Installed models on top, recommended next (with the why), everything clickable, on-disk sizes visible, delete where you see the file, drag a GGUF in to import it. Downloads confirm their size before starting — never surprise-download gigabytes.
• The task router. With several models installed, Quenderin suggests the best one per task — a tap to accept, never a silent switch. The routing logic is identical on every platform, enforced by machine-checked parity tests.
• The design became ours. Every color in the app and on this site is sampled from the elf artwork. The app icon, the in-app avatar, the website — one identity, teal and copper.

Earlier

The foundation: on-device inference via llama.cpp on macOS, iOS, and Android; the shared model catalog with RAM-fit gating and SHA-256 download verification; the tool-using agent with its hard safety blocklist; onboarding; and the measured performance numbers on the reality page.

← Back to Quenderin · Get help · Source on GitHub