I don’t pay for Perplexity or ChatGPT after combining my local LLMs with Perplexica and SearXNG

Published Aug 23, 2026, 7:30 AM EDT Ayush Pande is a PC hardware and gaming writer. When he's not working on a new article, you can find him with his head stuck inside a PC or tinkering with a server operating system. Besides computing, his interests include spending hours in long RPGs, yelling at his friends in co-op games, and practicing guitar. When it comes to simple inference tasks, my locally-hosted large language models are reliable enough to deliver accurate results. But everything falls apart as soon as I start prompting my LLMs about complex errors, new topics, or frequently updated information. By default, the models I run on llama.cpp can only access their training data, meaning they’re completely isolated from real-time information from the Internet. As such, even my bulkiest mixture-of-experts models are bound to provide outdated results if I tried using them without any MCP servers, APIs, or custom agent code that could connect them to the Internet. Fortunately, I ran into Perplexica a while ago, and this self-hosted alternative to Perplexity combines my local models with its own search engine, thereby delivering precise, up-to-date results while making my privacy sessions fairly private. Perplexica is a neat self-hosted alternative to Perplexity It even has SearXNG built into its container image If you’ve read my articles on XDA, you’ve probably seen me gush over SearXNG, and I have good reasons for relying on it instead of a single search engine. For one, SearXNG is a self-hosted metasearch engine that sends my queries through Google, Bing, and dozens of other search engines, thereby providing a diverse, unbiased set of answers. On the privacy front, it gets rid of cookies, tracking elements, and personal information. Unlike typical search engines, it also doesn’t create profiles to sell targeted advertisements. If anything, it also gets rid of ads and sponsored content from the results. For finer control over my query results, I can freely select which search engines SearXNG funnels my queries through. The best part? The official Perplexica container image deploys a local SearXNG server automatically, and it uses this metasearch engine to fetch real-time results for my LLM queries. As such, it fulfills the same task as Perplexity, ChatGPT, and other cloud platforms for free without restricting the number of searches I can perform or storing my data on some external company’s servers. The Perplexica web UI is pretty reminiscent of Perplexity’s interface, and it even lets me choose the available sources for my inference tasks. I also adore how Perplexica lists the references used when generating responses, so I can always check the sources instead of blindly following hallucinated, out-of-date information that LLMs are infamous for generating. Getting a Docker-based Perplexica instance up and running is just as easy, as you can execute docker run -d -p 3000:3000 -v vane-data:/home/vane/data --name vane --restart unless-stopped itzcrazykns1337/vane:latest within the Linux terminal. Technically, Perplexica supports a bunch of inference engines, but I rely on good ol’ llama.cpp, as it has minimal performance overhead and includes enough flags and tweaks to help me fine-tune my LLMs’ capabilities. While on the subject of my AI models… I usually cycle through certain models when running queries on Perplexica llama.cpp + Qwen3.6-35B-A3B work well for complex research tasks Seeing as how different LLMs have their own strengths and weaknesses, it shouldn’t come as a surprise that I host a couple of different models for my Perplexica-powered inference needs. For coding, troubleshooting, or intricate research queries, nothing beats Qwen3.6-35B-A3B’s reasoning prowess. And thanks to its mixture-of-experts design, I can host this 35B LLM on my outdated RTX 3080 Ti while maintaining token generation speeds in the 24 t/s range even with a context length of 200,000 tokens. Just to test things out, I tossed some errors I encountered randomly on Proxmox 9 and asked my Perplexica instance to help me fix them. In every case, Qwen3.6-35B-A3B (Q4_K_M) managed to pinpoint the source accurately and provided detailed steps to help me fix these problems. Likewise, I asked it for help with some ESP32 automation projects, and Perplexica was able to sift through the right forums, documentation pages, and guides to answer my queries. As for the search modes, I mostly stick to the Speed and Balanced profiles, as they’re accurate enough for the most part. That said, I must admit that using Qwen3.6 doesn’t result in immediate results, as I typically have to wait 2–3 minutes on the Speed (and almost twice as much on the Balanced profile) before Perplexica starts responding with the answer. Meanwhile, Gemma-4-E4B is pretty decent for simple queries Leaving aside tasks where I want accuracy above all else, my basic search queries don’t need something as bulky as Qwen3.6, which is where the Gemma-4-E4B (Q4_K_M) model running on my GTX 1080-powered llama.cpp server comes into the equation. That way, I can leave Qwen3.6-35B-A3B (and the PC running this LLM) engaged in my VS Code and Pi (the agent harness, I mean) tasks while running Gemma-4-E4B for general-purpose queries. Besides these transformation models, I also keep nomic-embed-text-1.5 running on my Pascal. After all, Perplexica needs an embedding model for RAG-aided tasks, and nomic-embed-text is light enough to run on my old hardware alongside a Gemma-4-E4B instance. Perplexica aids my research tasks without invading my privacy or charging regular fees As someone who avoids cloud platforms like the plague, Perplexica turned out to be a handy addition to my self-hosted LLM-powered productivity suite. After all, its ability to combine SearXNG’s live results with my locally-hosted LLM stack ensures I get the latest information from my prompts. I also appreciate that it labels the AI-generated responses with the proper references, so I can go back and verify dodgy-sounding information myself instead of blindly trusting my LLMs.