Now supporting 4 foundation models

Deploy AI Models in Minutes. Not Hours.

GPU Router removes infrastructure complexity and lets developers deploy AI models to GPU infrastructure with just a few clicks.

gpurouter deploy glm-4-9b-chat --size small
Live

Real-time AWS Spot capacity

AWS Spot Placement Score — 0 (unlikely) to 10 (very likely to get capacity right now).

Live score history is still building — check back soon.

Everything you need to serve a model

No EC2 wrangling, no security groups, no server flags. Just the parts that matter.

One-click deploy

Pick a model and a GPU tier — we boot the instance, load the model, and hand you a URL and API key.

GPU that fits

We match each model to the smallest GPU that can serve it, from a single A10G to an 8× H200 node.

OpenAI-compatible API

Every endpoint speaks the standard chat-completions API. Point any HTTP client at the URL.

TTL cost guards

Set an auto-terminate window so dev instances never quietly rack up a bill overnight.

Private by default

Endpoints are protected by an API key and can be locked to a CIDR range you control.

Built-in load testing

Measure TTFT and throughput from both client and server before you ship to production.

How it works

From zero to a running endpoint in four steps.

01

Pick a model

Choose from the catalog. Each entry shows the GPU tiers it runs on and what it costs.

02

Choose a GPU tier

Small, medium, or large. We size the instance, disk, and server flags for you.

03

Deploy

We boot the instance, load the model, and run a health check until it is ready.

04

Call your endpoint

Get a URL + API key. Point any OpenAI-compatible client at it and start shipping.

Supported models

Four production-ready foundation models, each pre-tuned for the GPUs it runs on.

View all models

GLM-4 9B Chat

Zhipu AI

Popular

A capable 9B BF16 dense chat model that runs natively on any Ampere-or-newer GPU — no FP8 tensor cores required.

Min GPU
A10G (24GB)
Memory
24 GB
Params
9B
Precision
BF16
Chat128K

Qwen 3.5 4B

Alibaba Qwen

Trending

A compact vision-language model that pairs strong reasoning with image understanding — light enough for a single 24GB GPU.

Min GPU
A10G (24GB)
Memory
24 GB
Params
4B
Precision
BF16
Vision-Language128K

Qwen 3.6 35B-A3B

Alibaba Qwen

New

An FP8 mixture-of-experts model with 35B total parameters — high quality with efficient FP8 serving on Ada/Hopper GPUs.

Min GPU
L40S (48GB)
Memory
48 GB
Params
35B (A3B MoE)
Precision
FP8
Reasoning128K

GLM-5.2

Zhipu AI

Frontier

A frontier-scale FP8 MoE model with sparse attention and 1M native context. Runs on an 8× H200 node.

Min GPU
8× H200
Memory
1128 GB
Params
756 GB MoE
Precision
FP8
Reasoning1M (capped 32K)

Why GPU Router

Built for people who want a private GPU-backed endpoint without hand-rolling infrastructure.

  • No EC2, security groups, drivers, or vLLM flags to hand-roll
  • GPU auto-matched to each model — pay only for what fits
  • State lives in AWS tags, so endpoint keys are never lost
  • One command up, one command down — billing stops instantly
  • Server-side metrics scraped for real throughput numbers
  • Restrict access to a CIDR range and rotate keys anytime
We went from a blank AWS account to a live inference endpoint before our standup finished. It is genuinely a few clicks.
AC
Ava Chen
ML Engineer, Northwind
The GPU-fit sizing alone saved us from over-provisioning. The load-test report is exactly the data I needed.
MR
Marcus Reid
Founding Engineer, Loopwork
The CLI was already great. The dashboard makes it something I can hand to the rest of the team without a walkthrough.
PN
Priya Nair
Head of AI, Fathom

Frequently asked questions

Everything you might want to know before deploying.

No. GPU Router picks a GPU that fits your model, boots the instance, loads the model, and returns a ready-to-call URL and API key. When you are done, one action tears it down and stops billing.

Ship your first endpoint today

Deploy a foundation model to your own GPU in minutes. No infra code, no console clicking.