{
  "id": 17816431,
  "title": "My phone runs a local 2B model that searches the web, and Perplexity feels pointless now",
  "url": "https://urgent.news/2026/11/02/my-phone-runs-a-local-2b-model-that-searches-the-web-and-perplexity",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-02T15:30:22.000Z",
  "source": {
    "name": "XDA Developers",
    "slug": "xda-developers",
    "url": "https://www.xda-developers.com/phone-runs-local-model-that-searches-the-web-perplexity-feels-pointless/"
  },
  "original_language": "en",
  "account": "Running a local AI model on a phone has become possible for a couple of years, yet most people still prefer cloud AI tools on their devices. The reason being, powerful models on mobile are usually limited to cloud servers. However, a decent local LLM with web search functionality on mobile can deliver good results without compromising accuracy or depth. This specific function is still experimental but exists.\n\nPerplexity, a search engine among the bunch, can be compared to such apps. Its answers are usually tidy little reports with headers and sources included. Despite having paid for ChatGPT, Gemini, and Claude, the user doesn't appreciate having three of them. Perplexity's free plan offers unlimited basic searches, with Pro Search limited to roughly five a day. However, the user isn't keen on adding another subscription for quick lookups that likely don't need the Pro models.\n\nThe owner of Perplexity abandoned its ad program in February 2026, the same month ChatGPT started showing ads on its free and cheapest tiers. Both decisions were made by people whose priorities differ from the user's. Perplexity also sells shopping integrations and enterprise search on top of subscriptions, making the plain lookup a small part of what the company is building.\n\nPrivacy is an obvious argument for using local AI models. However, it's not the only factor. For this setup to work, a local LLM app with a web function and a decent model that can handle tool calling are needed. Tiniest models struggle with recognizing when a question needs a search and calling the tool correctly. For this, a model of at least 2B is required.\n\nThe user uses Noema on iOS and Off Grid on Android for their local LLM needs. Both apps have web search functionality and check whether the model supports tool calling. LLM Hub, which covers both Android and iOS, is recommended if Noema isn't an option. Android also has Local LLM, where Gemma 4 models can search the web after confirmation.\n\nThe user's preferred setup is Qwen 3.5 2B at Q4_K_M in Noema on an iPhone 16. Web search is a single toggle in the tools menu, and searches go through Nooma Search, a SearXNG instance that doesn't require an account or attach an identifier to queries. The model only reads short passages of text when a web search is used, making responses faster. The user keeps Reasoning/Thinking off to avoid rambling and wasting tokens. With a 2B model, the phone's 8GB is left free for other uses. Quantization preserves reasoning and holds onto the knowledge signal with less confidence, making 4-bit models a good choice.",
  "summary": "A small model that has web is a powerful model",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}