OPENJET HOSTED INFERENCE
OpenJet.
On a GPU of your own.
A dedicated machine running Qwen3.8-27B with our CUDA inference engine. We handle the initial setup. You connect your agent or application.
THE OFFER
Hosted inference.
Set up for your work.
Start with a dedicated RTX 3090, 24 GB VRAM, and Qwen3.8-27B Q4_K_M. Choose your region when you order. Availability is confirmed before payment.
Per dedicated RTX 3090 with OpenJet and Qwen. Setup and payment handled directly by email.
Order by email ↗- OpenJet engine and model installed
- A dedicated inference machine for your workload
- Endpoint connection details and a working configuration
- A setup check with your client or agent
- One active inference sequence per GPU
Manual provisioning today. We confirm context limits, access arrangements, billing and support scope by email.
RUN THE NUMBERS
What will your inference cost?
Compare GPU rental prices and the cost per million generated tokens.
Published mean: code, explanation, writing and migration · thinking off
Published RTX 3090 / Qwen3.8-27B Q4_K_M measurements. Results and settings ↗
Prices, calculation and benchmark conditions
Rental token cost = hourly price × 1,000,000 ÷ (generation tok/s × 3,600). *Runpod uses the published RTX 3090 llama.cpp benchmark as an estimate; its UD-Q4_K_XL template differs from the benchmark’s Q4_K_M. Vast offer #55480654 is one RTX PRO 6000 with 96 GB VRAM in Beijing at $1.152/hour, plus bandwidth. Its token-cost estimate uses the published Qwen3.8-27B NVFP4 (4-bit) vLLM result: 128.33 tok/s at concurrency 1, MTP depth 3, one GPU. NVFP4 differs from GGUF Q4_K_M. This is a matching GPU-family benchmark reference, not a measurement of offer #55480654; the supplied listing does not specify the card’s exact edition or power limit. The model must be configured for NVFP4 and MTP to match this reference. Bandwidth is excluded from the token-cost estimate. RTX PRO 6000 benchmark and configuration ↗
The $0.30/hour OpenJet hosted rate, $0.50/hour Runpod template and $1.152/hour Vast offer are the recorded prices used for this comparison. Available deployments and their prices can change. Vast uses the supplied offer #55480654; bandwidth is additional.
Workflow speeds are the published measurements, with no projected uplift. Reasoning uses the HTTP result; short prompts use four drafts and thinking off. File-context rates come from the published context sweep, which retained the best prefill and decode rates independently. Quantization, context and draft settings differ between deployments. OpenJet’s representative generation rate is the published four-prompt mean of 106.7 tok/s, not the Python prompt’s maximum. Token-cost bars use hourly rental prices and the stated generation throughput. One active inference sequence runs per OpenJet GPU. Full benchmark methodology ↗
PRICES & CONFIGURATIONS
Vast.ai ↗Recorded deployment price · 13 Oct 2026
Offer #55480654 · 1× RTX PRO 6000 · 96 GB · Beijing, CN · bandwidth extra
Runpod · Qwen deployment ↗Recorded deployment price · 13 Oct 2026
RTX 3090 · llama.cpp · UD-Q4_K_XL · 192K context
LET’S SET IT UP
Tell us what
you want to run.
Order by email. We confirm availability and payment details, then set up your machine and send the connection details.
- 01 Send your requirements
- 02 Arrange payment
- 03 Receive your connection details
Before you order.
How do I order and pay?
Email your requirements to order an RTX 3090 running Qwen3.8-27B with the OpenJet engine. Setup and payment are handled directly by email. The comparison shows provider compute costs; storage and network charges are extra.
Do I pay when the model is idle?
A dedicated machine reserves compute for you. Plan for all running hours, including time between requests. Agree how the machine is stopped, restarted and billed before setup; there is no automatic usage billing or stop control on this page.
How fast is it?
Published RTX 3090 measurements include 78 tok/s for reasoning over HTTP and a 106.7 tok/s mean across four greedy generation prompts. These use different workloads and are not hosted service guarantees. We check your configuration during setup. Read the methodology and results ↗
Can multiple people share one machine?
The current OpenJet engine runs one active inference sequence per GPU. A shared machine needs requests queued or separate capacity. It does not provide simultaneous independent sessions on one GPU.
Can I choose UK or EU hosting?
You can request a region. We confirm the available location and hosting arrangements before payment. Any specific residency, security or compliance requirements need to be discussed and verified for that deployment.
Can I still run OpenJet for free?
Yes. The software is open source under AGPL-3.0. Install it on your own hardware or arrange your own cloud machine. This paid offer covers a hosted machine and agreed setup work.