What’s new in Quo

Last updated 3 November 2026. Also on GitHub.

Notable changes to Quo. Quo Web and the desktop app ship together under the desktop version (desktop-v* tags); the phone app has its own numbers. The format follows Keep a Changelog.

Quo Desktop 1.5.0

2026-10-21

Added

• Any model from Hugging Face — the desktop app can search Hugging Face and run almost any GGUF model (about 160 kinds: Llama, Qwen, Gemma, Mistral, DeepSeek, GLM, Phi, Granite and more). It says before downloading if a model won't work, and suggests a size that fits your computer.
• A faster engine for NVIDIA graphics cards — Models & storage offers llama.cpp's CUDA engine where it helps: long messages are read about 40% faster and replies written about 10% faster. It's downloaded only if you ask, from llama.cpp's own releases, and checked before use; if it ever fails to start, Quo goes back to the engine it came with.
• Linux on ARM — Quo now has ARM builds for Linux.
• Qwen3.8 27B replaces Qwen3.5 27B in the desktop app: the newer model, with the same speed-up for writing several words at once. If you have the old one, Models & storage offers to remove it.
• Gemma 4 models — Gemma 4 E2B in the browser (Chrome, Edge and Firefox; best with a graphics card), and Gemma 4 12B and 26B A4B in the desktop app. Google's models write well in 140+ languages, read images in the desktop app, and can think when you turn it on.
• Dark mode and themes — follows your system, or choose Light or Dark in Settings → Appearance. Dark is true black by default (clean on OLED screens), or Midnight, the indigo night with frosted glass. The accent is Indigo or Mono (greys, no glow). In the desktop app the window and the approval window follow the same choice.

Fixed

• No engine from a CDN — in Safari and older browsers, the offline model's engine came from a public CDN (jsDelivr). Quo now ships that version itself and loads it only where it's needed.

Improved

• Faster writing in the desktop app — Qwen and Gemma models draft several words at a time and check each one, so replies are the same, only quicker: rewriting a file with a small change went from about 100 to 570 words a second in our test. Gemma 4 models come with a small drafter file for this.
• Longer memory on smaller computers — when memory is short, the desktop app stores the model's working memory in 8 bits, which fits about twice as much conversation for about 3% of speed.
• Much faster follow-ups with offline models — the model now reuses what it already read of the chat instead of reading it all again: the prompt no longer changes every minute (it has the date, not the time), and long chats and Code sessions drop old parts in steps. In a test in the browser, a follow-up in a 2,300-token chat took 2 seconds instead of 114.
• Clearer replies — every model is asked to write plainly: the answer first, short sentences, no filler, and the same style in the code and text it writes.
• Code colours — code in replies and in the desktop file panel is coloured by language (about 50, with the grammars VS Code uses), in calm colours that follow your theme and accent. Code blocks are light in Light mode, like GitHub.
• Maths — formulas are drawn properly, inline and on their own line, however the model writes them; prices like $5 stay as text.
• Easier tables and text — numbers line up and right-align, headings and paragraphs wrap without a lone last word.
• Small details — the background glow fades out through a fine dither (Light and Midnight), changing the look dissolves the new one in, terminal output matches the dark style, and text selection, the text cursor and scrollbars take the accent.
• Easier to read — small grey text and green labels pass contrast checks in both themes; sheets stay readable where the browser can't blur.

Quo Desktop 1.4.0

2026-10-20

Added

• Push and pull requests — Code mode can push the branch you're on and give you a link to open a pull request on GitHub. You approve every push; it never force-pushes. Turn it off in Models & storage → Code.
• Searches and sources — each search shows as a step while it runs, and the pages a reply used appear as sources under it.
• The model decides when to search — no Web switch. Quo's cloud models, the free one included, search through OpenRouter when a question needs it (3 searches a day on the free model). Settings → Web says whether models may search at all.
• Code mode starts itself — the model asks for a project folder when a request needs one (you approve it), and can run a dev server in the background, hand a broad question to a read-only helper, keep a checklist, and summarise long sessions.
• Thinking levels — Off, Quick, Auto or Deep, showing only the levels each model has. Auto leaves it to the model.
• Approvals in Quo's own window, a file panel with clickable file links in replies, and opening what Quo built in your browser.
• Memory meter and summaries — see how full the model's memory is and summarise earlier messages to keep going. Type / for commands and @ to point at a project file.
• Connection awareness — an Offline pill, and a choice to move to your offline model when a cloud model can't be reached.

Improved

• Faster offline Qwen3.5 in the desktop app (editions with a prediction head: about twice as fast on code edits). Models that don't fit memory start with a smaller window instead of failing, and an idle model unloads after 30 minutes.
• Model settings follow each maker's advice (sampling, window, reply length). Qwen3.5 0.8B and 2B left the desktop app as too small; their files can be deleted in Models & storage.
• Commits match the project's message style, a failing git hook comes back for the model to fix, signed commits work, and commands that throw work away carry a warning.
• Quo's co-author line names the quo-code account, so GitHub shows "quo-code" with Quo's picture instead of "quo-ai[bot]".

Fixed

• macOS: Code mode refuses system folders reached through links (/etc is /private/etc).
• Linux: the taskbar shows Quo's icon instead of a generic one.

Quo Desktop 1.3.0

2026-10-18

The first releases of Quo Web and the desktop app.

Added

• Quo Web — offline models in the browser (WebGPU in Chrome and Edge), saved chats, thinking, resumable downloads, installable as an app.
• Quo for desktop — macOS, Windows and Linux. Models run in llama.cpp on the graphics card: hundreds of words a second, bigger models (up to Qwen3.5 27B), photos, and models you already have in LM Studio or Ollama.
• Code mode — a coding agent for a project folder: it reads, edits, runs tests and commits, and you approve every change, command and commit. Works with cloud models, your own key or an offline model.
• Your own API keys — OpenRouter, Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, Groq, or your own server (Ollama, LM Studio). Messages go straight to the provider.
• Web search in Code mode and with your own keys or offline models: Brave, Tavily, Exa, SearXNG or DuckDuckGo.
• Cloud models — a free model, and frontier models (Claude, GPT, Gemini and more) that unlock with credits.
• Photos, documents, voice and read aloud, all processed on your computer.
• Memories — things Quo should know about you, read by offline models only.
• Quo Sync — end-to-end encrypted sync of chats and memories across web, desktop and Android, with sign out of all devices and account deletion.
• Open source — Apache 2.0, with a contributing guide, code of conduct and security policy.

Quo Phone 1.5.0

2026-10-04

Added

• Gemma 4 — new model with image understanding
• Memories — Quo remembers facts you ask it to, like your name
• Light and dark mode — refreshed look

Improved

• Answers — smarter, more accurate and more on topic
• Speed — faster replies; thinking mode no longer takes so long
• Smaller download — the starter model installs in the background (Play Asset Delivery fast-follow)

Fixed

• A crash when using thinking mode, plus many stability fixes

Quo Phone 1.4.0

2026-07-07

Added

• Bundled starter model (instant start) — Qwen3.5 0.8B ships with the app via Play Asset Delivery and is materialized on first launch, eliminating the first-run download wall. Builds without the pack fall back to the download flow automatically
• Rolling context compression — Older messages fold into an on-device LLM-written summary once history approaches the context limit, so long conversations keep their early content instead of silently forgetting it
• Enhanced Recall (premium) — Optional 84MB embedding model indexes conversations on-device; when a question relates to summarized-away messages, the most relevant originals are retrieved verbatim into the prompt
• Read Aloud / offline voice output (premium) — Optional ~535MB OuteTTS + WavTokenizer voice reads any answer aloud, fully offline, from the message long-press menu
• Voice input trial — Free users get 3 metered voice transcriptions before the paywall instead of a hard feature wall
• New models — Gemma 3 4B (vision), Gemma 3 1B, Phi-4 Mini (3.8B reasoning), Llama 3.2 3B
• Speed Boost (experimental) — Draft-model speculative decoding: the 0.8B starter drafts for Qwen 2B/4B when enabled in Settings → Intelligence
• Storage management — Download/delete controls for the optional voice and recall models in Settings
• Suggested questions — Empty state now shows 4 random questions from a 1,100+ knowledge bank instead of workflow template buttons

Fixed

• Responses cutting off mid-sentence — Generation now auto-continues (up to 3 extra passes) when the token cap is hit, so detailed multi-step answers run to completion
• Free vision contradiction — Users on non-vision models are pointed to the free Ministral vision model instead of being shown a paywall for a capability the free tier includes
• Conversation storage scalability — Conversations now store one-per-key with an index (plus one-time migration), avoiding Android CursorWindow limits that could make large histories unreadable
• Partial answers no longer discarded — If generation errors after streaming began, the already-streamed content is kept instead of thrown away

Improved

• Smooth streaming with live formatting — Replies render through the full markdown + LaTeX pipeline while streaming, with typewriter pacing and automatic repair of in-progress markdown (bold/code/fences/math no longer flash as raw markers)
• Background memory usage — Optional TTS and recall AI contexts are released when the app goes to background and reload lazily
• Dependencies — llama.rn 0.12.5, whisper.rn 0.5.5, lucide 1.x, expo-audio replaces deprecated expo-av, IAP 14.7.20; removed 15 unused dependencies
• Engineering — Jest test suite introduced (95+ unit tests), all model download URLs verified, deduplicated generation pipeline shared between send and resend

Quo Phone 1.3.2

2026-07-06

Added

• New and refined model options, including premium models
• Voice input auto-detects the language with multilingual models

Improved

• Vision model downloads and model management

Fixed

• Layout issues, bugs and performance problems

Quo Phone 1.3.0

2026-03-30

Fixed

• Native Crash (SIGSEGV): endCompletion() race condition — Added isCompleting mutex to prevent stopCompletion() from being called when no completion is active, fixing use-after-free crashes in librnllama on dotprod and v8 devices
• Native Crash (SIGSEGV): JSI binding lambda — unloadModel() now waits for any in-flight completion to finish before releasing the native context, preventing the JSI lambda from accessing freed memory
• ANR: Model loading blocks UI thread — Deferred initLlama() to InteractionManager.runAfterInteractions with a 100ms yield so the heavy native memory-mapping no longer competes with input dispatching during app startup
• ANR: Download completion blocks main thread — Moved moveAsync() file I/O in the download .done() callback off the native main thread via setTimeout(0), fixing "RNBackgroundDownloaderModuleImpl.onSuccessfulDownload — Input dispatching timed out"
• ANR: Streaming bridge contention — Increased token streaming throttle from 50ms to 100ms to halve the number of React Native bridge crossings and reduce setState pressure on the UI thread
• ANR: Mali GPU shader compilation — The InteractionManager deferral also prevents glLinkProgram from contending with input dispatching on Mali GPU devices during model init
• Math/LaTeX rendering overhaul — Completely rewrote math rendering to convert LaTeX to Unicode inline within text, fixing broken paragraph flow. Supports Greek letters, super/subscripts (Unicode), fractions, roots, operators, relations, set theory, calculus, logic, arrows, accents, delimiters, trig functions, and environments. Display math renders as code blocks; inline math as code spans

Added

• File attachments — Attach text files, PDFs, code, CSV, JSON, XML, and more to messages. The app extracts text content and sends it to the model so you can ask questions about documents, analyse finances, review code, etc. Supported via the new 📎 button in the input bar

Improved

• Chat scroll during streaming — Scrolling up to read earlier content no longer fights auto-scroll. A floating ↓ button appears to jump back to the latest output. Auto-scroll resumes when you scroll back to the bottom or when generation finishes
• Model tags updated — Qwen3.5 models now carry the "Recommended" badge; legacy Qwen3 models are marked "Outdated" with descriptions pointing to their newer replacements

---

Quo Phone 1.2.9

2026-03-04

Added

• Background Generation: Model responses now continue generating when the app goes to background, with a persistent notification so Android keeps the process alive
• Generation Notifications: Get notified when your AI response finishes while the app is in background, with a preview of the response
• Expanded Question Database: Added 585+ diverse, timeless knowledge questions across 30 categories (science, history, philosophy, technology, cultures, etc.) for better conversation starters
• Copy Responses: Users can now copy model responses to clipboard (already implemented)

Improved

• Model Knowledge Confidence: Updated system prompts to answer factual questions confidently instead of hedging or deflecting on common knowledge
• Android Compatibility: Fixed deprecated API warnings for Android 15+ by removing explicit statusBarColor setting
• Large Screen Support: Enabled all orientations and removed portrait lock to better support tablets and foldable devices
• Model Management: Improved custom model deletion to properly unload and remove files without affecting other models
• Prompt Quality: Removed weak/trivial questions and upgraded all prompts to be more specific and engaging

Fixed

• Background Generation Bug: Previously, tapping home button during generation would stop the response and discard progress. Now generation continues in background
• Model Deletion Bug: Fixed issue where deleting a model could unload unrelated active models
• System Prompt Issues: Fixed model tendency to deflect on basic science questions by clarifying when to use "I don't know"
• Notification Cleanup: Fixed edge cases where ongoing generation notifications could persist after completion

Technical

• Added persistent notification channel "Generation In Progress" for Android
• Added proper notification dismissal across all generation lifecycle events (stop, error, completion, return to app)
• Enhanced AppState handling to track background/foreground transitions correctly
• Improved error handling for model context invalidation scenarios

---

Quo Phone 1.2.8

• Initial release with core AI chat functionality
• On-device model inference
• Vision/image analysis capabilities
• Memory system for user preferences
• Multi-language support