llm-launcher

Open-weights LLMs on GPUs billed by the minute.

Pick a model and a GPU from a live marketplace. Get an OpenAI-compatible endpoint, a chat and a coding workspace on it.

Launch a model

Browse live GPU offers without an account.

qwen3.8-27b-fp81× RTX A6000 · 48 GB total VRAM · Sweden, SERunning
GPU rate$0.40/hrUsage to date $0.35GPU $0.33 · Network $0.00 · Storage $0.02
Context window32.8k tokens

Traffic

Tokens
48.2k31.0k in · 17.2k out
Prompt cache
62%19.2k input tokens reused
Speed
41 tok/s128 requests

GPU load

Now
34%
Average
41%
VRAM now
44.1 GBof 48.0 GB
Account billed
$0.35
Provider total
$0.30
Commission
15%
ComponentUsageProvider rateProvider costAccount billed
GPU0.750 hours$0.380/hour$0.29$0.33
Storage0.750 hours$0.020/hour$0.02$0.02
Download30.9 GB$0.00/GB$0.00$0.00

6m 12s

  1. Reserve4s
  2. Rent3s
  3. Boot1m 08s
  4. Connect22s
  5. Download2m 41s
  6. Start engine1m 54s
  7. Ready

+0sEndpoint ready

Up to 32 users · Smooth up to 1 user

UsersFeelsFirst token p50 / p90Speed per userTotal output
1Smooth0.2 / 0.3 s38 tok/s38 tok/s
16Overloaded60 / 115 s38 tok/s38 tok/s
32Overloaded124 / 236 s38 tok/s38 tok/s
Base URL
https://inference.llm-launcher.com/v1
Model
qwen3.8-27b-fp8
API key
Exported as LLM_SAAS_API_KEY

Check it with curl

curl https://inference.llm-launcher.com/v1/chat/completions \
  -H "Authorization: Bearer $LLM_SAAS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3.8-27b-fp8","messages":[{"role":"user","content":"Hello"}]}'

Pay for minutes, not hours.

From a few cents an hour. Provider rate plus 15%, per started minute. Prepaid, no subscription.

Vast.ai median
GPUVRAMPer hour
RTX 306012 GB$0.07
RTX 309024 GB$0.19
RTX A600048 GB$0.40
RTX 409024 GB$0.51
L40S48 GB$0.64
RTX 509032 GB$0.73
A100 SXM480 GB$0.80
H100 SXM80 GB$2.93
H200141 GB$4.34
B200192 GB$8.13

Available offers

  • 1× RTX 306012 GB VRAM per GPU$0.052 / hourfree model download12 GB total VRAMPLAMD Ryzen 5 5600X · 12 cores32 GB RAM↓ 940 · ↑ 860 Mb/sTraffic not billed by this host
  • 1× RTX 309024 GB VRAM per GPU$0.130 / hour+ $0.004 model download24 GB total VRAMDEAMD Ryzen 9 5950X · 32 cores64 GB RAM↓ 2,400 · ↑ 1,800 Mb/sTraffic at $1.00 per TB
  • 1× B200192 GB VRAM per GPU$7.50 / hourfree model download192 GB total VRAMUSIntel Xeon Platinum 8570 · 56 cores512 GB RAM↓ 9,800 · ↑ 9,400 Mb/sTraffic not billed by this host
  • 1× RTX 409024 GB VRAM per GPU$0.290 / hour+ $0.017 model download24 GB total VRAMSEAMD EPYC 7543 · 32 cores128 GB RAM↓ 3,100 · ↑ 2,900 Mb/sTraffic at $4.00 per TB
  • 2× RTX 309024 GB VRAM per GPU$0.270 / hourfree model download48 GB total VRAMROIntel Core i9-12900K · 24 cores64 GB RAM↓ 910 · ↑ 810 Mb/sTraffic not billed by this host
  • 1× H100 SXM80 GB VRAM per GPU$1.85 / hourfree model download80 GB total VRAMCAAMD EPYC 9354 · 64 cores256 GB RAM↓ 8,600 · ↑ 8,100 Mb/sTraffic not billed by this host
  • 1× RTX A600048 GB VRAM per GPU$0.380 / hour+ $0.037 model download48 GB total VRAMFIAMD EPYC 7302 · 32 cores128 GB RAM↓ 1,900 · ↑ 1,700 Mb/sTraffic at $9.00 per TB
  • 1× H200141 GB VRAM per GPU$2.64 / hourfree model download141 GB total VRAMUSIntel Xeon Platinum 8468 · 48 cores384 GB RAM↓ 9,200 · ↑ 8,800 Mb/sTraffic not billed by this host
  • 1× RTX 509032 GB VRAM per GPU$0.410 / hour+ $0.004 model download32 GB total VRAMFRAMD Ryzen 9 9950X · 32 cores96 GB RAM↓ 2,000 · ↑ 1,900 Mb/sTraffic at $1.00 per TB
  • 1× A100 SXM480 GB VRAM per GPU$0.620 / hour+ $0.066 model download80 GB total VRAMNLAMD EPYC 7763 · 64 cores256 GB RAM↓ 5,400 · ↑ 5,100 Mb/sTraffic at $16.00 per TB
  • 1× L40S48 GB VRAM per GPU$0.520 / hour+ $0.017 model download48 GB total VRAMJPIntel Xeon Gold 6430 · 32 cores192 GB RAM↓ 4,200 · ↑ 3,900 Mb/sTraffic at $4.00 per TB

Use your model your way.

In the browser, or inside the tools you already use.

Chat

Open WebUI, signed in as you, one click from a running machine.

Open WebUI answering a SQL question with the qwen3.8-27b-fp8 model

Code

OpenCode in the browser, wired to the model you launched, editing its workspace.

OpenCode adding a retry with backoff to a TypeScript function, shown as a diff

Connect

Your model speaks the OpenAI API. Plug it into your app, a local OpenCode, an agent framework or a script: point it at one URL, keep your code.

app.py
from openai import OpenAIimport osclient = OpenAI( base_url="https://inference.llm-launcher.com/v1", api_key=os.environ["LLM_SAAS_API_KEY"],)reply = client.chat.completions.create( model="qwen3.8-27b-fp8", messages=[{"role": "user", "content": "Triage this ticket"}],)
~/.config/opencode/opencode.json
{ "model": "llm-saas/qwen3.8-27b-fp8", "provider": { "llm-saas": { "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "https://inference.llm-launcher.com/v1", "apiKey": "{env:LLM_SAAS_API_KEY}" }, "models": { "qwen3.8-27b-fp8": {} } } }}

Frequently asked questions.

When does billing start and stop?

Billing starts when the machine is rented, so boot and model download are included, and stops when you destroy it. Every started minute is billed. A launch that fails is billed only for the minutes the machine was rented.

Which models can I run?

A curated catalog of open-weights Qwen models, from a 2B that fits a 12 GB card to a 27B in BF16 or FP8. Each one is sized against the GPU memory it needs, so you only see machines that fit.

Which GPUs can I rent?

Any GPU on a verified Vast.ai host with compute capability 8.0 or newer and a recent CUDA driver: from an RTX 3060 to the RTX 4090 and 5090, the A100, L40S, H100, H200 and B200. The launch page only lists the machines whose memory fits the model you pick.

Is the API compatible with OpenAI?

Yes. Every model is served by vLLM behind an OpenAI-compatible /v1 endpoint. Point an OpenAI SDK, OpenCode or any client that accepts a base URL at it, with an API key from your account.

How long does a launch take?

Usually a few minutes. The launch page shows each stage as it happens (reserve, rent, boot, connect, download, start engine) and you can destroy the machine at any of them.

What happens when my credit runs out?

Credit is prepaid and loaded by card. A machine whose owner has no credit left is stopped automatically, so your balance never goes negative.

Do I need an account to look around?

No. Live GPU offers are open to everyone. Launching a model needs a free account and some credit.

Where do the machines come from?

From the Vast.ai marketplace. You choose the exact offer; it shows the provider's hourly rate, and a 15% commission is added to what you are billed.

Pick a GPU. Launch a model.

Launch a model Contact us