A 4 MB antidote to algorithmic enslavement

What it does

Zhihu++ is a third-party Android client for Chinese Q&A platform Zhihu, built to strip out ads, sponsored content, paid columns, and the official app’s algorithmic manipulation. It weighs under 4 MB versus Zhihu’s 110 MB+ official client, and supports full browsing, search, commenting, notifications, and content export without the bloat.

The interesting bit

The project treats recommendation as a user-rights issue, not merely a convenience. Its “local recommendation algorithm” runs entirely on-device via web scraping, letting users customize weights rather than accepting Zhihu’s engagement-optimized feed. A Full variant bundles ONNX for offline LLM-embedding-based content filtering—vector-similarity NLP blocking that never phones home. The author is admirably frank that this is partly a “technical attempt” with undeveloped ambitions like local knowledge bases.

Key highlights

• Feed autonomy: Web, Android, local, and mixed recommendation modes; toggle between logged-in and anonymous states to break filter bubbles
• Aggressive filtering: Regex blocklists, per-user/per-topic blocks, paid-content suppression, and (in Full) AI semantic blocking via local embeddings
• Content portability: Export answers and entire collections as PDF, Markdown, HTML, or images; save pictures without watermarks
• API flexibility: Implements Zhihu’s zse96 v2 signing to call 99% of web APIs, plus Android API emulation for completeness
• Minimal telemetry: Optional and granular—launch counts, timestamps, IP, and SHA256-hashed user IDs only; browsing history and algorithm data stay local

Caveats

• The author has unsubscribed from issues and PRs citing burnout, so community contributions may face delays
• Full version’s AI features are experimental; the author notes “a lot of functions are not yet developed”
• Project self-describes as “not yet perfect”—expect rough edges in a rapidly evolving reverse-engineered client

Verdict

Worth a look if you’re a Zhihu user who resents platform control over your attention, or curious about on-device ML replacing cloud curation. Skip it if you need polished enterprise support or don’t read Chinese content; this is a principled indie project, not a product.