Summarize YouTube videos, web articles, and pasted text on your phone. Paste a link (or share one into the app) and get skimmable notes, a detailed prose view, or a bulleted outline.
Status: working. YouTube transcript summarization, web-article and pasted-text summarization, Groq cloud, on-device inference, saved sessions, resumable checkpoints, and source Q&A are functional as of version 1.5.0.
Sonario has two engines, and you pick which to use per summary:
- Groq cloud (recommended) - sends your text to Groq's API and summarizes with Qwen 3.6 27B. Sonario splits large sources into rate-safe requests, waits for Groq's minute windows, and checkpoints each completed call. You bring your own API key. Your text goes to Groq's servers.
- On-device - runs a downloaded GGUF locally through llama.cpp. Source text remains private after Sonario fetches it, but phone inference is slower than Groq and can warm the device during long summaries.
- Summarizes YouTube captions, web articles, pasted text, PDF, EPUB, DOCX, TXT, and Markdown files.
- Uses
qwen/qwen3.6-27bas the only Groq cloud model, preventing retired model IDs from being restored by an old setting or saved session. - Queues cloud requests against conservative TPM/RPM limits and Groq's live reset headers instead of repeatedly failing with minute-based 429 errors.
- Offers three newer local choices with clear quality/speed tradeoffs: Qwen3 4B, Gemma 3n E4B, and LFM2 2.6B.
- Keeps long cloud summaries alive with a foreground service and retries temporary DNS, timeout, and network-handoff failures.
- Saves up to 12 recent sessions locally, restores the latest session on launch, and resumes from checkpoints without repeating completed model calls.
- Includes Normal, Detailed, Bullets, and chapter views, plus source-grounded Ask about this Q&A.
- Lets users delete one session or clear all locally stored transcripts, summaries, checkpoints, and Q&A history.
Left: paste a link and pick an engine. Center: a summarized YouTube video with Normal / Detailed / Bullets views. Right: first-run setup.
- Install the APK (see below) and open Sonario.
- On the first screen, tap Use Groq cloud instead to skip the local-model download.
- Get a Groq API key at console.groq.com and create a key.
- In Sonario's Settings, paste the key and tap Save key.
- Back on the main screen, make sure the toggle is on Groq cloud, paste a YouTube link or article URL, and tap Summarize.
Qwen 3.6 27B's published Groq free-tier baseline is 8K tokens per minute and 200K tokens per day, applied across the whole Groq organization. Sonario uses slightly lower internal working limits for safety, follows the provider's live remaining-token/reset headers, and displays a countdown between calls. Extra keys in the same organization share the same quota and do not multiply it.
- Open Sonario and choose a local model:
- Qwen3 4B Instruct 2507 for the strongest overall local summaries and Q&A.
- Gemma 3n E4B Instruct for nuanced long-form summaries and writing.
- LFM2 2.6B for the smallest download and fastest local responses.
- Tap Get and keep the phone on Wi-Fi. Downloads range from about 1.6 GB to 4.3 GB.
- When it finishes, select On-device, paste a source, and tap Summarize.
YouTube links in any common shape all work:
youtu.be/VIDEO_ID(short link)youtube.com/watch?v=VIDEO_IDm.youtube.com/watch?v=VIDEO_ID(mobile)youtube.com/shorts/VIDEO_ID,/live/VIDEO_ID,/embed/VIDEO_ID- Extra share parameters like
?si=...are ignored.
You can also share a link from the YouTube or browser app straight into Sonario via the Android share sheet. Web article URLs and raw pasted text also work.
The two engines differ, and the app shows which is active:
- On-device: summarization happens locally. The only network calls are the one-time model download and fetching whatever link you paste.
- Groq cloud: the text to summarize is sent to Groq. Don't use this for anything you wouldn't send to a third-party API. Your Groq key is stored only on your device and is sent solely to Groq.
Sonario uses a layered extractor rather than relying on one caption URL:
- Load the watch page in one cookie-preserving session and handle YouTube's consent page when it appears.
- Read the current InnerTube API key, WEB client version, visitor data, and the
real
getTranscriptEndpoint.paramsvalue from YouTube's page data. - Prefer
youtubei/v1/next->youtubei/v1/get_transcript, which returns the transcript segments directly as JSON. - Fall back to consistently identified ANDROID and WEB
playerrequests, then download an available caption track. - Parse JSON3, legacy XML, SRV3, TTML, or WebVTT captions.
The extractor preserves signed caption URL parameters when changing formats and reports Proof-of-Origin-token tracks accurately instead of claiming the video has no captions. These are still undocumented YouTube endpoints, so a future YouTube change can require another extractor update. Failed requests show Extractor build 2 diagnostics so you can confirm the new APK is actually installed.
- The Groq cloud path is pinned to
qwen/qwen3.6-27b; old model preferences and saved Scout session IDs can no longer control the request model. - Source chunks, detailed-output budgets, chapter excerpts, and Ask excerpts are sized so one call fits beneath the free-tier 8K TPM ceiling.
- Sonario leaves headroom below the published TPM/RPM/daily limits, reads Groq's
x-ratelimit-remaining-tokensand reset headers, and waits with a visible countdown before the next call. - Routine summaries use Qwen's non-thinking mode to avoid spending output tokens on hidden reasoning that is unnecessary for summarization.
- When Groq reports daily exhaustion, Sonario stops rather than waiting all day. Every completed call remains checkpointed so Resume continues later.
- Cloud requests retry transient DNS, Wi-Fi/mobile-data handoff, connection, and
timeout failures for up to ten minutes instead of immediately ending with
Unable to resolve host api.groq.com. - A foreground service holds a partial CPU wake lock and a temporary Wi-Fi lock only while a summary or source question is active. The notification displays rate-limit waits and network-retry status.
- Groq responses are buffered before being committed to a summary, so a failed connection can be retried without duplicating a partial response.
- The Ask box shows the actual API/network error in place, keeps the typed question after a failure, and searches relevant passages across the whole source instead of sending only the first portion of a long video or book.
- Long jobs have a visible Cancel control, stale errors clear when a new source is entered, and the app no longer treats a failed Ask call as a successful generic answer.
| Model | Download | Best for | Tradeoff |
|---|---|---|---|
| Qwen3 4B Instruct 2507 Q4_K_M | ~2.5 GB | Best overall local summaries, comprehension, and follow-up answers | Slower than LFM2 |
| Gemma 3n E4B Instruct Q4_K_M | ~4.24 GB | Nuanced long-form summaries and polished writing | Largest and slowest option |
| LFM2 2.6B Q4_K_M | ~1.56 GB | Fast summaries, extraction, and lower storage use | Weaker on difficult or subtle material |
Sonario downloads the GGUF files from public Hugging Face repositories into
filesDir/models, Android's private app-specific storage. The app checks free
space before starting, supports resumable downloads, and removes obsolete models
from versions 1.4 and earlier.
Android deletes app-specific files when Sonario is uninstalled. Sonario also sets
hasFragileUserData=false and excludes models/ from cloud backup and
phone-to-phone transfer, so multi-gigabyte model files are not retained or
restored after reinstall.
The app uses Llamatik 1.8.1, an Android wrapper around llama.cpp. It requests available acceleration but may fall back to CPU depending on the device and the runtime build. Local models remain useful for privacy and offline work, while the 27B Groq model remains substantially stronger and faster for everyday use.
If the app ever crashes, it writes the stack trace to its private app files and shows it on the next launch with a Copy button instead of silently closing.
Kotlin/Jetpack Compose. Both engines implement one interface, InferenceEngine,
and the summarize pipeline talks only to that:
llm/InferenceEngine.kt- shared interface (ensureReady,stream).llm/LlmEngine.kt- on-device via Llamatik/llama.cpp.llm/GroqEngine.kt- Qwen 3.6 through Groq's OpenAI-compatible streaming API.llm/RateLimiter.kt- local pacing plus synchronization with Groq reset headers.llm/ModelDownloader.kt- resumable GGUF downloads with free-space checks.data/Settings.kt- engine choice and Groq key (local preferences).source/SourceFetcher.kt- YouTube (InnerTube) and web-article fetching.summarize/SummarizeEngine.kt- map-reduce summarizer with bounded cloud calls and small, phone-safe local chunks.summarize/Prompts.kt- prompts carried over from Sonario desktop.ui/- Compose screens, theme, settings, system meter, and crash screen.CrashReporter.kt- global uncaught-exception logger.
To add another backend, implement InferenceEngine and hand it to
SummarizeEngine. GroqEngine is the template for an HTTP-based engine.
See BUILD.md.
Port of Sonario desktop by pgotta. On-device inference via llama.cpp/Llamatik; cloud inference via Groq.
MIT. See LICENSE.
The Recent sessions card has a Clear button. After confirmation, it permanently deletes all locally saved session folders, including source transcripts, chapter data, summaries, checkpoints, and saved Q&A. Exported files outside the app are left alone.
Sonario saves summaries locally instead of keeping the only copy in an Activity/ViewModel. The most recent session is restored after an app or Activity restart, and a Recent sessions panel can open, resume, or delete prior work.
For long summaries, Sonario checkpoints after every completed LLM call. If Android or the network interrupts the run, Resume skips already-completed condensed chunks and derived views so those Groq tokens are not spent twice. Completed source text and Ask history are stored with the session. Up to 12 recent sessions are retained in the app's private files directory.
The Ask field follows the software keyboard into view, supports two to four lines of editing, offers the keyboard Send action, and preserves unfinished text across ordinary Activity recreation.