Inference rental · launch price

Qwen3.8 27B on your own RTX 5090,
$0.70 an hour.

One dedicated GPU, vLLM, NVFP4 weights with multi-token prediction. An OpenAI-compatible /v1 endpoint and a web chat panel, deployed in one click. One tenant, billed per hour, hosted in Poland.

$0.70/h
287.4 tok/s decode
$0.68 / 1M generated
single tenant
Poland · EU

See the plans

Our price vs the Qwen API USD per 1M tokens

	Alibaba API, official	Ours, RTX 5090 $0.70/h*	You save*
Output, 1M tokens	$3.00	$0.68	77%
Input, 1M tokens	$0.50	$0.011–$0.032	94–98%
Real workload: 1M output + 20.3M input	$13.16	$0.90–$1.34	90–93%

* Estimate if the 5090 is fully occupied: $0.70/h ÷ tokens it serves in that hour.
Sources: API price — Alibaba Model Studio pricing, qwen3.8-27b model page · 5090 throughput 287.4 tok/s decode, 6,000–18,000 tok/s prefill — vLLM NVFP4 MTP@3 benchmark · workload mix (81.7M output, 1.66B input) — our run data

Plans hours × $0.70

Hourly

1 h minimum

$0.70 net / 1 h

Decode: 1,034,640 tok max

Day

24 h

$16.80 net / 24 h

Decode: 24,831,360 tok max

Week

168 h

$117.60 net / 168 h

Decode: 173,819,520 tok max

Specs what you get

GPU: NVIDIA RTX 5090, 32 GB
Model: Qwen3.8-27B
Quant: NVFP4 weights · FP8 KV cache
Serving: vLLM · MTP@3 · 4 streams
Endpoint: OpenAI-compatible /v1 · web chat
Tenancy: One tenant, whole GPU
Deploy: One click
Location: Poland · EU