Verdict: Vane, the project formerly known as Perplexica, wins for privacy, price, and anyone who already runs their own hardware. Perplexity wins for zero-setup answer quality. If you have a Docker host and a 7B-or-larger model served by Ollama, Vane gives you a cited, Perplexity-style answer engine for nothing per month, and your queries never leave your network. If you have no hardware and no appetite for a Compose file, Perplexity's free tier plus Pro at $20/month ($17/month billed annually, per perplexity.ai/pro) is the sensible purchase. Vane is not a full replacement: Deep Research, the Comet browser, and Perplexity's hosted Computer agent have no equivalent in it.
feat(app): rename to 'vane' on 9 March 2026 (SHA 39c0f19) in the itzcrazykns1337/perplexica to itzcrazykns1337/vane. Docker Hub still hosts both names, which is why most tutorials still work.
Perplexica is an open-source answer engine: you ask a question, it runs a web search through a bundled SearxNG instance, reads the results, and writes an answer with inline citations. It was built as an open alternative to Perplexity and released under the MIT licence.
In March 2026 the maintainer renamed the project to Vane and widened the framing from "Perplexity clone" to a general-purpose answering engine. The repository is the primary record: the rename commit is dated 9 March 2026, container paths moved from /home/perplexica to /home/vane, and the canonical image tag is now itzcrazykns1337/vane:latest. Search volume still sits on the old name, so if a guide tells you to pull itzcrazykns1337/perplexica, it is describing the same software at an older name. Both image names still resolve on Docker Hub.
Per the project README as extracted on 15 September 2026, Vane covers the core answer-engine loop and a fair amount around it:
SEARXNG_API_URL at your own instance.
What it does not do: frontier-model quality out of the box, Perplexity's Deep Research reports, the Comet browser, or the hosted Computer agent that runs multi-step tasks on Perplexity's infrastructure. If those are the features you actually use, self-hosting is not a swap.
Perplexity's pricing is published: Pro at $20/month or $17/month on an annual plan, Max at $200/month or $167/month annually, and Education Pro at $10/month with verification. The free plan gives extended access to Pro Search under weekly limits, which the help centre's plan table describes as around three Pro Searches per day.
Vane's licence cost is zero. The real bill is hardware and electricity, plus any cloud API keys you attach. That is the honest comparison: Perplexity charges a predictable subscription and absorbs the compute; Vane moves the compute onto your machine, where the cost is lumpy up front and small thereafter. For a single user with an existing GPU workstation, Vane is cheaper within weeks. For a team that would otherwise buy one GPU per person, it usually is not.
The scale gap is worth naming. Perplexity finalised a funding round at roughly a $20 billion valuation in September 2025 (Reuters, via Yahoo Finance). Vane is one maintainer and a contributor community. That asymmetry shows up in polish and in support, not in whether the answers carry citations.
This is where most self-hosting attempts fail. Reviews published in April 2026 put the practical floor around 14B parameters: below that, citation quality is noticeably shallow, because the model summarises the first source rather than reconciling several. At 14B and above it holds up for moderate research work. A reasonable hardware ladder from the same reporting: 8 GB of VRAM runs llama3.1:8b or mistral:7b; 16 to 24 GB runs qwen2.5:14b; 48 GB and up runs 70B-class models. nomic-embed-text is the unglamorous, correct choice for embeddings.
Two setup notes that save an afternoon. Ollama must listen on 0.0.0.0, not 127.0.0.1, or the container cannot reach it. That is the single most common failure report. And versions from 1.10.0 onwards support keep_alive, which holds the model in memory and removes the cold-start on the first query after idle. Expect around fifteen minutes for a Docker Compose install if your model runtime is already working. If you are still choosing a runtime, our comparison of Ollama and LM Studio covers that decision, and our local coding model benchmarks cover which weights to pull.
There is a broader reason the wrapper now matters more than the model. In our own testing: across three trials each on an identical seven-constraint article-planning task, Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence. Median wall time was 23 seconds for Gemini against 67 seconds for Opus (n=6, measured 15 September 2026). When configurable models match frontier models on structured work, the differentiator shifts to retrieval quality, source control, and where your data goes.
Privacy is the concrete argument, and it got more concrete in 2026. A class action filed in federal court in San Francisco on 1 April 2026 alleges that tracking code in Perplexity's product sent user chat information to Google and Meta before Perplexity itself received it. That is an allegation, untested at the time of writing, and it should be read as one. But it illustrates the structural point: with a hosted answer engine, your queries transit somebody else's infrastructure and you cannot audit what rides along. With Vane, the query goes to your SearxNG instance and your model.
The second argument is control. You choose the sources, the search depth, and the model per query, and nothing changes underneath you on a vendor's release schedule. Perplexity has been shipping fast, including agent products that have leaked ahead of announcement, as with the Teammate coding agent. Fast shipping is good until it changes behaviour you had built a workflow around.
Pick Vane if you already run local models, handle queries you would rather not send to a vendor, or want an answer engine you can point at academic sources and internal notes. Pick Perplexity Pro if your time is worth more than $20/month, you want frontier-model answers with no configuration, or you need Deep Research and the agent tooling. Running both is a legitimate answer: Vane for routine and sensitive lookups, Perplexity for the hard research sessions. If you are assembling a broader self-hosted stack, our roundup of free and open-source AI tools and our local voice cloning comparison cover adjacent pieces.
Q: Is Perplexica the same thing as Vane?
A: Yes. The project was renamed from Perplexica to Vane in a commit dated 9 March 2026, and the old GitHub URL redirects to github.com/ItzCrazyKns/Vane. Documentation written before then uses the Perplexica name for the same software.
Q: Do old Perplexica Docker commands still work?
A: Mostly. Docker Hub still hosts both itzcrazykns1337/perplexica and itzcrazykns1337/vane, so older pull commands resolve. New installs should use the vane image, and container paths are now under /home/vane rather than /home/perplexica.
Q: Can Vane replace a Perplexity Pro subscription?
A: For everyday cited web search, yes, provided you run a 14B-class model or larger. It does not replace Deep Research, the Comet browser, or the hosted Computer agent, so check whether you actually use those before cancelling.
Q: What hardware do I need to run Vane well?
A: Reviews from April 2026 suggest 8 GB of VRAM for 7B and 8B models, 16 to 24 GB for a 14B model such as qwen2.5:14b, and 48 GB or more for 70B-class weights. The 14B tier is where citation quality stops feeling thin.
Q: Does Vane require local models, or can it use cloud APIs?
A: Either. The README lists Ollama and LM Studio for local serving alongside OpenAI, Anthropic, Google Gemini, Groq, DeepSeek and AIMLAPI. You can also mix, using a local chat model with a hosted embedding model.
Q: Why does my Vane container fail to reach Ollama?
A: Almost always because Ollama is bound to 127.0.0.1. Set it to listen on 0.0.0.0 so the container can reach the host, then restart both services.
No corrections issued. Facts verified against the Vane repository README and Perplexity's published pricing on 15 September 2026. Pricing and repository figures change; check both primary sources before acting on the cost comparison.