vLLM
Researched 63d agoInference RuntimeTier 3vllm-project
vLLM is a self-hosted runtime, not a hosted model catalog. You bring compatible open-weight models, so there is no per-token pricing to track.
Covers 0 workload areas across 0 tracked models; last verified 2026-06-29.
Use it for
- Getting oriented before committing to a specific model
Do not use it for
- Final benchmark picks without opening the relevant model detail page
Tracked models
0
Models available through this provider
Priced output routes
0
Output pricing not yet tracked
Cheapest output
Unknown
Output pricing not yet tracked
Batch-ready models
0
No batch pricing tracked
Latest model release
Unknown
Release date of the newest tracked model
Freshness
2026-06-29
Researched 63d ago
Information
vLLM is a high-throughput, memory-efficient open-source inference engine and serving library for large language models. It is self-hosted rather than a managed token-priced API, and it ships an OpenAI-compatible HTTP server for local or private deployments. The vllm-project also maintains vLLM-Omni, a sibling runtime for serving omnimodal workloads.
Where this host wins
Not enough capability or benchmark coverage yet to call strengths for this provider.
Getting started
SDKs & libraries
Platform Overview
vLLM does not offer a fixed hosted model catalog. Operators bring their own open-weight Hugging Face model weights, run vLLM on their own compute, and serve compatible architectures through a local or private OpenAI-compatible endpoint. vLLM-Omni (github.com/vllm-project/vllm-omni) follows the same runtime pattern for omnimodal models across text, image, audio, video, and robotics inputs or outputs; the models it serves remain separate model-family and researcher decisions. LLMReference intentionally does not attach token-priced modelProvider rows to vLLM because the runtime has no per-model commercial offer or published token pricing; users pay their own GPU or infrastructure costs.