Which platform runs on a small VPS?
Short answer
anythingllm — AnythingLLM holds 394 MB at peak and runs one container, but it failed to ingest five of eleven documents. FastGPT is the lightest platform that finished most of the corpus, at 3.2 GB.
What this use case actually needs
- The host is a small VPS — two to four gigabytes of RAM, shared with other services
- Container count matters as much as memory, because each one carries its own supervision and restart cost
- The document set is modest; a few hundred files rather than a continuous feed
How the candidates measured up
| Platform | Documents ingested | Parsing | Retrieval | Peak memory | Ready |
|---|---|---|---|---|---|
| anythingllm | 6/11 | 100% | 41% | 394.2 MB | 146 s |
| openwebui | 11/11 | 61% | 62% | 912.9 MB | 18 s |
| fastgpt | 8/11 | 97% | 41% | 3,227 MB | 239 s |
| dify | 11/11 | 82% | 66% | 2,572.7 MB | 39 s |
Every figure comes from the same run against the same documents, so these are directly comparable. The full per-document and per-query rows are in the raw data.
The honest shape of this answer
On pure footprint AnythingLLM wins by a wide margin — 394 MB peak, one container, ready in 146 seconds. Nothing else measured comes close. Open WebUI is second at 913 MB, also one container.
But footprint is only half the question, and the other half goes against it: AnythingLLM ingested six of eleven documents. It embeds each document synchronously inside the upload call, and on the larger files that call did not finish. Its 100% parsing score covers only what it managed to take in, which is why the tables here mark it.
So the recommendation splits by how large your documents are:
- Small documents on a small host — AnythingLLM, comfortably
- Mixed or large documents on a small host — FastGPT at 3.2 GB, the lightest platform that got through most of the corpus, or Dify at 2.6 GB if you can spare the container count
What the memory range actually looks like
| Platform | Containers | Peak memory | Documents ingested |
|---|---|---|---|
| AnythingLLM | 1 | 394 MB | 6/11 |
| Open WebUI | 1 | 913 MB | 11/11 |
| Dify | 15 | 2,573 MB | 11/11 |
| FastGPT | 13 | 3,227 MB | 8/11 |
| RAGFlow | 5 | 10,326 MB | 11/11 |
The spread is twenty-six-fold, and container count does not predict it. RAGFlow runs the fewest containers of the three purpose-built platforms and needs the most memory by a factor of four, because it bundles Elasticsearch. A requirements page that quotes one of these two numbers without the other is not telling you much.
Open WebUI deserves a second look here
At 913 MB and one container it costs roughly twice AnythingLLM, and unlike AnythingLLM it ingested the entire corpus and retrieved 62% of the answers. If your VPS can spare the extra half-gigabyte, it is the better small-host choice on every measure except raw footprint.
Where this recommendation stops applying
- Your corpus contains large HTML documents — AnythingLLM failed on five of eleven, including every SEC filing
- The instance will be reachable from a network; AnythingLLM ships with authentication disabled and hands out full-access API keys to unauthenticated callers
- You need visibility into how documents were chunked; it exposes no way to list stored chunks