Self-hosted (vLLM)

Run open weights on your own GPUs behind an OpenAI-shaped API

docs.vllm.ai github.com/vllm-project/vllm

Overall in llm apis

10of 18 · score 63.2

Data coverage

85%

23 of 27 fields sourced

Best for

•Data that legally cannot leave your infrastructure
•Sustained high-utilisation inference where GPU-hours beat per-token billing
•Guaranteeing a model version survives indefinitely

Strong on

+Zero retention Yes 1/18
+AU region Yes 1/18

Weak on

−Batch discount 0% 12/14
Better: Anthropic: 50%
−Effort control No 12/18
Better: Anthropic: Yes
−Audio in/out No 11/18
Better: OpenAI: Partial
−Tool use Partial 13/18
Better: Anthropic: Yes

vs Cohere

•Ahead on Zero retention: Yes vs Partial
•Ahead on AU region: Yes vs Partial
•Behind on Tool use: Partial vs Yes
•Behind on Image input: Partial vs Yes

How Self-hosted (vLLM) scores among 18 llm apis23 of 27 fields known

Capability7

Flagship model

Whatever you deploy, DeepSeek-V3.x, Qwen3, Kimi K2, Llama 4

· 2026-01-15

Context window

—

Max output

—

Image input

Partial

10/18

Leader: Anthropic · Yes

· 2026-01-15

Audio in/out

11/18

Leader: OpenAI · Partial

· 2026-01-15

Open weights1st

Yes

1/18

· 2026-01-15

OpenAI-compat API1st

Yes

1/18

docs.vllm.ai · 2026-01-15

Performance2

TTFT p50

—

Output tok/s

—

Pricing4

$/M input

N/A

raw.githubusercontent.com · 2026-11-07

$/M output

N/A

raw.githubusercontent.com · 2026-11-07

$/M cache read

N/A

raw.githubusercontent.com · 2026-11-07

Batch discount

12/14

Leader: Anthropic · 50%

Agentic6

Tool use

Partial

13/18

Leader: Anthropic · Yes

· 2026-01-15

Schema output1st

Yes

1/18

docs.vllm.ai · 2026-01-15

Effort control

12/18

Leader: Anthropic · Yes

· 2026-01-15

Computer use

5/18

Leader: Anthropic · Yes

· 2026-01-15

MCP support

6/18

Leader: Anthropic · Yes

· 2026-01-15

Cache TTL

Automatic prefix cache in VRAM, no TTL

docs.vllm.ai · 2026-01-15

Governance & continuity6

Continuity policy

Permanent, and this is the whole point of the baseline. The checkpoint sits on your disk. Nobody can deprecate it, re-price it, silently upgrade it, or change its behaviour between releases. The obligation transfers to you: you own the GPUs, the upgrades, the security patches and the on-call roster.

· 2026-01-15

Notice period

N/A

raw.githubusercontent.com · 2026-11-07

Pinnable versions1st

Yes

1/18

· 2026-01-15

Zero retention1st

Yes

1/18

· 2026-01-15