Perplexica with SearXNG and local LLMs instead of Perplexity and ChatGPT
The writer doesn't pay for Perplexity or ChatGPT. Perplexica runs in Docker with its own SearXNG, and local models on llama.cpp answer with live sources. A 35B mixture-of-experts model on an RTX 3080 Ti handles hard troubleshooting and research, while a small Gemma model on an old GTX 1080 takes simple queries, with nomic-embed-text for embeddings.
Setup steps
- Run the container: docker run -d -p 3000:3000 -v vane-data:/home/vane/data --name vane --restart unless-stopped itzcrazykns1337/vane:latest
- Serve a large model with llama.cpp for hard queries (they used Qwen3.6-35B-A3B, Q4_K_M)
- Serve a small model on a second GPU for simple queries (Gemma-4-E4B)
- Run nomic-embed-text for embeddings
- Connect the models in Perplexica and use Speed or Balanced mode
Steps summarized from the original Blog source ↗. Use “Copy” to grab them with the tools and source link, or copy an AI prompt.
How it went
The writer gave it real Proxmox 9 errors and says the big Qwen model found the source and gave fix steps every time; it also found the right forums and docs for ESP32 projects. The 35B model runs at about 24 tokens a second with a 200,000-token context on an RTX 3080 Ti.
Worth knowing
It isn't instant: expect 2–3 minutes before an answer on Speed mode with the big model, and about twice that on Balanced. Use a small model for simple lookups.
Only what the original source ↗ says, re-read Nov 2, 2026. How we write these
Try it with AI
Paste this into ChatGPT, Claude or your own agent for step-by-step help. Prompt written by AppToBot from this recipe's data, not by the original author. It keeps drafts for your review and asks before anything is sent or paid.
Help me set up this automation using Perplexica/Vane (Docker, SearXNG built in) + llama.cpp with Qwen3.6-35B-A3B and Gemma-4-E4B: "Perplexica with SearXNG and local LLMs instead of Perplexity and ChatGPT". Goal: replace Perplexity Pro and ChatGPT Plus. How it should work: - Trigger: You ask a question - Then: SearXNG fetches live results - Then: A local model on llama.cpp reads them - Then: It writes an answer with references - Result: An up-to-date answer with no subscription Setup steps (summarized from the original source): 1. Run the container: docker run -d -p 3000:3000 -v vane-data:/home/vane/data --name vane --restart unless-stopped itzcrazykns1337/vane:latest 2. Serve a large model with llama.cpp for hard queries (they used Qwen3.6-35B-A3B, Q4_K_M) 3. Serve a small model on a second GPU for simple queries (Gemma-4-E4B) 4. Run nomic-embed-text for embeddings 5. Connect the models in Perplexica and use Speed or Balanced mode Guide me one step at a time. Ask me for anything you need (accounts, sheet or channel names) before you start. Safety rules: - Build and test with sample data first, and show me a sample result before anything goes live. - Keep anything that goes to other people (emails, posts, replies, record updates) as drafts for my review until I say otherwise. - Never send, post, buy, pay, delete or change records without my approval. - Tell me about any cost (paid plans, API usage) before I'd incur it, and help me set a spending limit. - Keep API keys and passwords out of shared docs and chats. - Privacy: Queries and answers stay on your machines apart from the searches SearXNG sends; SearXNG strips cookies and tracking (per writer). Based on: https://www.xda-developers.com/i-combined-perplexica-and-searxng-with-local-llms/
Tool names link to each official site (opens in a new tab).
Running cost: Free to run Open-source software on GPUs the writer already owns. Self-hosted How we sort costs
Replace an app: AI bots that replace Perplexity
App vs bot, side by side
| Old app / habit | Bot | |
|---|---|---|
| Tool | Perplexity Pro, ChatGPT Plus | Perplexica + llama.cpp on two home GPUs |
| Monthly cost | Not reported | Free on hardware the writer already owns |
| Setup | None (already in use) | Not reported · Advanced |
| Your time | Baseline | Not reported |
| Maintenance | Handled by vendor/you as before | Run and update the llama.cpp servers and the container yourself. |
| Privacy | Data stays with the app vendor | Queries and answers stay on your machines apart from the searches SearXNG sends; SearXNG strips cookies and tracking (per writer). |
Full comparison
| Replaces | Perplexity Pro, ChatGPT Plus |
|---|---|
| Bot / agent stack | Perplexica + llama.cpp on two home GPUs |
| Monthly cost before | Not reported |
| Monthly cost after | Free on hardware the writer already owns |
| Setup time | Not reported |
| Time saved | Not reported |
| Skill level | Advanced |
| Reliability | 3/5: firsthand, with specific models, hardware and wait times; slow, and the hard queries need a serious GPU. |
| Maintenance | Run and update the llama.cpp servers and the container yourself. |
| Privacy & data | Queries and answers stay on your machines apart from the searches SearXNG sends; SearXNG strips cookies and tracking (per writer). |
| Best for | Home-lab users with spare GPUs |
| Not for | Anyone who wants instant answers |
Pros and cons
Pros
- No search limits or subscription
- Live results with references
- Big and small models split across GPUs
Cons
- 2–3 minutes before an answer on Speed mode with the big model, about twice that on Balanced
- Needs a GPU like an RTX 3080 Ti for the hard queries
Source and evidence
First-person XDA Developers article (Aug 2026): the writer tested it on real Proxmox 9 errors and ESP32 projects and says it found the cause in every case. No usage period.
Read the original Blog source ↗
“Not reported” means the source doesn't say. We never fill gaps with guesses. Numbers marked “per author” or “vendor claim” are the source's own and weren't independently tested. Source last checked Nov 2, 2026.
Our rating is an editorial opinion based on the public source above, not lab testing. Spot an error? Request a correction.
Plain-text summary (for search and AI assistants)
Perplexica with SearXNG and local LLMs instead of Perplexity and ChatGPT. Replaces Perplexity Pro, ChatGPT Plus with Perplexica + llama.cpp on two home GPUs. Cost before: Not reported. Cost after: Free on hardware the writer already owns. Setup: Not reported, Advanced. Time saved: Not reported. Reliability 3/5: 3/5: firsthand, with specific models, hardware and wait times; slow, and the hard queries need a serious GPU. Best for: Home-lab users with spare GPUs. Not for: Anyone who wants instant answers. How it went (per source): The writer gave it real Proxmox 9 errors and says the big Qwen model found the source and gave fix steps every time; it also found the right forums and docs for ESP32 projects. The 35B model runs at about 24 tokens a second with a 200,000-token context on an RTX 3080 Ti. Worth knowing: It isn't instant: expect 2–3 minutes before an answer on Speed mode with the big model, and about twice that on Balanced. Use a small model for simple lookups. Source: https://www.xda-developers.com/i-combined-perplexica-and-searxng-with-local-llms/
Related recipes
- Perplexity vs Google: an answer engine instead of search (3/5)
- AI job search agent with Claude for Chrome (3/5)
- A self-hosted Perplexity alternative with Vane and Ollama (3/5)
More: all Browsing recipes · AI agents that replace browsing and search · AI bots that replace Perplexity
Questions, corrections or ideas?
Spotted an error, want a recipe covered, or have a partnership idea? Send us a message and we'll get back to you.
Starting your own newsletter? Try beehiiv ↗
What is beehiiv?
beehiiv is a newsletter platform: you write and send emails, host a newsletter website, and grow and monetize a list from one place. We've used it for our own newsletter.
beehiiv says (vendor claims, not tested by us):
- Free plan with up to 2,500 subscribers and unlimited sends
- Built-in website, recommendation network and analytics
- 0% cut of paid subscription revenue (Stripe fees apply)
Good for: starting or growing an email newsletter. Not ideal for: email automations or removing beehiiv branding on a $0 budget (those need Lite or Pro).
Free plan: $0, no card needed. Paid plans from $49/mo billed annually (Lite, up to 1,000 subscribers). People who sign up through our link get a 14-day trial plus 20% off their first 3 months, per beehiiv's partner page. Facts as of Oct 2026, from beehiiv's own website. More about beehiiv →