Skip to content
State of GPU Compute 2026

Nexpod Hub

Open models, one deploy away.

Tested templates for language, image, speech and video models, plus fine-tuning kits. Change a few settings, pick a GPU, and run it on a server or behind an endpoint.
Browse the catalogPublish a template
Nexpod Hub: template cartridges for vLLM, ComfyUI, Axolotl and Jupyter on a staircase above an 8-GPU B200 server, the first one sliding into its slot.
hub / vllm-serve · v2.4 4/4 checks
  • Dockerfile6 lines
  • handler.py84 lines
  • hub.jsonmanifest
  • tests/
  • test_smoke.py2 tests
  • README.mddocs
Release checks
  • image build2m 10s
  • smoke · 1× H10038 s
  • smoke · 1× H20041 s
  • licence scanApache-2.0
$ nexctl hub fork vllm-serve

Start from code that runs

Every template ships with an image, a handler and tests that passed on real GPUs.

Deploy · diffusion-xl41 s
v3.2
diffusion-xl
template
live
dxl-endpoint
run.nexpod.ai
  • ✓image cached on node0.4 s
  • ✓weights from volume dxl-weights29 s
  • ✓worker w-1 healthy38 s
  • ●https://dxl.run.nexpod.ai live41 s

One click to an endpoint

Pick a template and get a live URL, billed per GPU-second.

dxl-endpoint · workers · 24 hpeak 13

Scales with use

A Hub endpoint follows traffic like any other: 1 worker overnight, 13 at the 15:30 peak, back down by midnight.

In plain words

What is Nexpod Hub?

A catalog of ready-to-run templates and open models for Nexpod. Start from one that works, fork it to change what you need, or publish your own so others can deploy it in a click.

How it works

Find it, set it, ship it.

Three steps from the catalog to a running model, in the console or with nexctl.
console.nexpod.ai/hub/llama-class-70b
official / llama-class-70b-instruct
Llama-class 70B Instruct
tests passed · v2.4
Precisionfp8bf16int4
MAX_BATCH64
GPU1× H200· €4.00/GPU·hr
Serverless endpointCloud Server
Deploy
  1. 01
    01 · browse

    Find a template

    Filter by modality, GPU and licence. Official and community entries side by side.

  2. 02
    02 · one click

    Set and deploy

    Change the exposed settings, pick a family and a target, press deploy.

  3. 03
    03 · contribute

    Fork and share

    Copy a template to your Git, improve it, and publish the new version back.

Catalog

Templates and open models, tested on Nexpod.

Each card lists the task and the hardware it was tested on. Deploy to an endpoint or a whole server.
› Summarise the contract in five bullets
› Which clause covers termination?
ttft 182 ms42 tok/s
official/llama-class-70b-instruct

Llama-class 70B Instruct

General chat and reasoning model, served with an OpenAI-style API and streaming.

Task
Text → Text
Tested on
1× H200 · fp8
Release
v2.4 · tests passed
Deploy to
Endpoint · Server
1024 × 1024 · 28 steps› a ridge at dusk, 35mm
official/diffusion-xl

Diffusion XL

Text-to-image at 1024 px, with LoRA loading and a batch queue for overnight runs.

Task
Text → Image
Tested on
1× H100
Release
v3.2 · tests passed
Deploy to
Endpoint · Server
Sample catalog: names are generic open-model classesSample

Community

Publish a template with one Git tag.

Add a manifest and a smoke test to a public repo and tag a release. Nexpod builds the image, runs the test on real GPUs and lists it after a short review.
Read the manifest spec
hub.json
{
  "name": "whisper-diarize",
  "version": "1.2.0",
  "task": "audio-to-text",
  "image": "./Dockerfile",
  "handler": "handler.py:run",
  "gpu": {
    "families": ["H100", "H200"],
    "count": 1
  },
  "settings": {
    "MODEL_ID": "whisper-class-large",
    "DIARIZE": true
  },
  "tests": ["tests/test_smoke.py"],
  "licence": "Apache-2.0"
}
release · acme/whisper-diarizev1.2.0
  1. tag v1.2.0 pushed12:01:07
  2. image built from Dockerfile2m 10s
  3. smoke test · 1× H10038 s
  4. smoke test · 1× H20041 s
  5. review by the Hub teamqueued
  6. listed under communityafter review
Listing preview
community / whisper-diarizeAudio → Text · 1× H100

FAQ

Hub, answered.

What the catalog is, how deploys work and how to publish.
What is Nexpod Hub?
A catalog of templates and open models that already run on Nexpod. Each entry has a container image, a request handler, settings you can change in a form, and hardware it was tested on.
Can I use Hub templates in production?
The Hub is in public beta. Templates marked official are maintained by Nexpod and versioned; community templates are tested on every release but supported by their authors.
Why start from a template instead of my own image?
You skip writing the Dockerfile, the handler and the health checks. A template exposes the settings that matter, like model ID, precision and batch size, and you can still fork it and change anything.
Who is the Hub for?
Developers who want a model behind an endpoint today, authors who want their open-source work to be one click away, and teams that want one vetted starting point instead of a wiki page.
How do I deploy something from the Hub?
Open a template, adjust its settings, pick a GPU family and choose a target: a Serverless endpoint or a Cloud Server. The deploy runs the same checks the template passed on release.
How do I publish my own template?
Add a hub manifest and a test file to a public Git repository, then tag a release. Nexpod builds the image, runs the tests on real GPUs and lists the template after a short review.

Customers

Teams shipping open models.

Product teams, labs and agencies that would rather start from a template that works than from an empty Dockerfile.
Placeholder

Placeholder quote. An AI app team says, in one or two plain sentences, what changed after moving production traffic to Nexpod.

Customer nameEngineering lead, AI app
  • Customer 01
  • Customer 02
  • Customer 03
  • Customer 04
  • Customer 05
  • Customer 06
  • Customer 07
  • Customer 08
Customer logosPlaceholder

Start now

Pick a model. Deploy it now.

Browse the catalog without an account. Deploying takes one, and the first GPU-second is the first thing you pay for.

Browse the catalogBook a demo
On-demand ratesEUR · excl. VAT
  • B300€6.00/GPU·hr
  • B200€5.00/GPU·hr
  • H200€4.00/GPU·hr
  • H100€3.00/GPU·hr