Open models,on a machine you rent

The Ollama server, with a chat page in front of it.The same port serves the API your own tools call.

Ollama 0.34.04 cores · 8 GBNVIDIA GPUServed on 8080

hub://ollama

$ dxflow workflow create --identity ollama hub://ollama --start --link

01Pick a machinecores, memory and a rate per hour

02Create the workflowpulled from the hub, once

03Run it, then stopstopped when the work is done

The application

Pick a model,and start talking

A chat page over the Ollama server, with the API alongside it.

A chat page

Pick an installed model and watch the answer stream as it comes.

And the API

The Ollama HTTP API sits at /api, behind that same one password.

The card comes with it

A GPU is in the definition, and Ollama loads as many layers into VRAM as fit.

How you run it

A password, a model,and the card

Who gets in, what is waiting when they do, and whether it answers on the GPU.

PASSWORDnginx checks it as basic auth for the user dxflow, across the whole port. It ships as dxflow, and everyone reading this knows that.

STARTUP_MODELPulled at start and selected in the UI. smollm2:135m is already inside the image, so the first start is quick.

resource.app.gpunvidia, so the models answer on the card. Override it to nothing and the same image runs them on the cores alone.

start it your way

$ dxflow workflow start ollama --override env.app.PASSWORD=something-long

$ dxflow workflow start ollama --override env.app.STARTUP_MODEL=llama3.1:8b --link

$ dxflow workflow start ollama --override resource.app.gpu= --link

The session

One port,two ways in

The chat page and the API share 8080 and one credential.

8080Webthe chat page

/apiAPIfor your own tools

/volumeStoragethe models you pull

Open it from anywhereStart with --link and the page comes back on an HTTPS address.

Set a password firstPASSWORD defaults to dxflow, and everyone can read this page.

Good to know

One port,one credential

The chat page and the HTTP API are the same door, and the same password opens both.

How the page reaches the server

A React app behind nginx proxies to Ollama on 11434. The UI calls it under /ollama/api, and the standard API sits at /api on the same 8080.

It finds the card on its own

Ollama offloads as many layers as fit into VRAM — nvidia-smi shows it while a model answers. Drop the gpu resource and it falls back to CPU without being told, which suits the small models.

Pull more whenever

ollama pull from a terminal adds models, and they land on the volume rather than back in the image.

The image

Pulled once,then it stays

Ollama arrives as one image. This is what comes down the first time, and what the disk should have free for it.

3.6Gamd64compressed, the way the registry counts it

2.8Garm64compressed, the way the registry counts it

50GOn diskunpacked, with room to work beside it

ghcr.io/dxflow-ai/ollama:latestPublished from the hub, pulled on the first start and kept for the ones after it.

Either architectureamd64 and arm64 are both published, and the machine pulls the one it runs.

What it asks for

What it wants,is what it needs

The definition asks for 4 cores and 8 GB, and the image wants all of it. A start given --fit still caps each step to whatever the machine actually has.

4 cores · 8 GBAsks forwhat the definition writes down

4 cores · 8 GBRuns onthe same figure, with nothing spare

NVIDIAGPUrequired, and filtered for

The ask is the floorThe definition and the image agree on this one, so a smaller machine has nothing left to trim.

--fit caps it to the hostA start given --fit trims each step to what the machine actually has, for that start alone. The definition is never rewritten.

Machines that fit it

Ollama asks for 4 cores and 8 GB, with an NVIDIA GPU. Cheapest first.

G4dn XLarge

$0.658/ hour4 cores · 16 GB · Tesla T4Start this machine

NC4as T4 v3

$0.658/ hour4 cores · 28 GB · Tesla T4Start this machine

G2 Standard-4

$0.884/ hour4 cores · 16 GB · NVIDIA L4Start this machine

Run Ollama on your own machinePick a machine that meets it, and it opens about a minute after you ask.

Start Ollama Browse workflows