Self-hosted (vLLM)

Self-hosted (vLLM) — Run open weights on your own GPUs behind an OpenAI-shaped API. The baseline every hosted option is measured against. vLLM serves open-weight models on your own hardware with an OpenAI-compatible endpoint, continuous batching, paged attention and prefix caching. You trade per-token billing for GPU-hour billing and an operational burden, and you get the only genuinely permanent answer to deprecation: the weights are on your disk and nobody can retire them. Founded 2023. Open source under Apache-2.0. We carry it as the do-it-yourself baseline, not as a vendor.

We track it in 1 comparison — LLM APIs — so every claim below is a cell in a table you can open and check rather than an impression. Across those rosters it sits against 17 other tools, and what follows is where it visibly separates from them.

Where it wins.

• Zero retention — Yes, which only 2 of 12 tools here manage.
• AU region — Yes, which only 4 of 17 tools here manage.

Each of those is ranked against the whole roster on its category page, not against a hand-picked subset, so a first place here means first of everything we list.

Where it gives ground.

• Batch discount — 0 %, 8th of 10. Anthropic records 50 %.
• Effort control — No. Anthropic records Yes.

None of these disqualify it on their own. They are the fields to check against your own requirements before you commit, because they are the ones where a competitor genuinely does better.

Provenance. 19 of 27 tracked fields carry a value for Self-hosted (vLLM), and 3 of those cite a document you can open. Last verified 2026-01-15. Every figure keeps its own provenance — measured by us, claimed by the vendor, inferred, or community-reported — and we would rather print a dash than a guess.

Its nearest neighbour in our data is Qwen. Self-hosted (vLLM) is ahead on Schema output (Yes against Partial) and Open weights (Yes against Partial). Qwen takes Batch discount (50 % against 0 %) and Trains on your data (Partial against No). That pattern repeats across the rest of the roster — see Self-hosted (vLLM) alternatives for the other rivals, each compared the same way.

Founded

2023

Licence

Apache-2.0 (open source)

Fields we track

19 of 27

Last verified

2026-01-15

The reference point rather than a recommendation. Self-hosting only beats a hosted API on cost at genuinely high, steady utilisation, and it costs you the agentic features — effort control, computer use, MCP — that make the frontier APIs worth their price. Choose it for sovereignty or permanence, not to save money.

Zero retention

●Yes1st of 12

Inferredverified 2026-01-15

Batch discount

0 %8th of 10

Inferred

AU region

●Yes1st of 17

Inferredverified 2026-01-15

Schema output

●Yes1st of 18

Vendor-claimedverified 2026-01-15source

Pinnable versions

●Yes1st of 18

Inferredverified 2026-01-15

Trains on your data

○No7th of 18

Inferredverified 2026-01-15

Effort control

○No12th of 18

Inferredverified 2026-01-15

OpenAI-compat API

●Yes1st of 18

Vendor-claimedverified 2026-01-15source

Self-hosted (vLLM) is the right call in these situations, each one drawn from a field we actually record:

• Data that legally cannot leave your infrastructure.
• Sustained high-utilisation inference where GPU-hours beat per-token billing.
• Guaranteeing a model version survives indefinitely.
• Zero retention is your binding constraint. Self-hosted (vLLM) records Yes, which only 2 of the 12 tools in the LLM APIs roster do. We define that field as whether prompts and completions can be excluded from all storage, including abuse-monitoring logs;.
• AU region is your binding constraint. Self-hosted (vLLM) records Yes, which only 4 of the 17 tools in the LLM APIs roster do. We define that field as whether inference can be pinned to Australian infrastructure;.

Reach for something else when any of the following is a requirement rather than a nice-to-have:

• Batch discount. Self-hosted (vLLM) records 0 %, 8th of 10 in the LLM APIs roster. Anthropic records 50 % on the same field. If that is a hard requirement rather than a preference, start elsewhere.
• Effort control. Self-hosted (vLLM) records No on the LLM APIs table. Anthropic records Yes on the same field. If that is a hard requirement rather than a preference, start elsewhere.

We publish this block because a comparison that only lists what a tool is good at is marketing. Every figure above sits on the same page as its source, and the field definitions are on the category tables if you want to check how we measured them.

What we don't know. 8 of the 27 fields we track for Self-hosted (vLLM) are still blank: Context window, Max output, TTFT p50, Output tok/s, $/M input, $/M output, $/M cache read and Notice period. Those render as dashes rather than as zeroes or assumptions, because an empty cell and a bad cell are not the same thing and only one of them is honest. If you know any of these figures and can point at a document, tell us.