pip install velar-sdk · $10 free credits

Your Python
function.
Any GPU.
In minutes.

No Dockerfile. No Kubernetes. No YAML. Decorate your function — Velar builds the container, provisions the GPU, and runs your workload at per-second cost.

Start for free Read the docs

$10 free GPU credits

Per-second billing

A100 from $1.74/hr

transcriber.py

import velar app = velar.App("transcriber") image = velar.Image.debian_slim().pip_install("faster-whisper") @app.function(gpu="L4", image=image)def transcribe(audio_url: str) -> str:

from faster_whisper import WhisperModel model = WhisperModel("large-v3") segments, _ = model.transcribe(audio_url) return " ".join(s.text for s in segments) # Transcribe 1000 files in parallel — runs on separate GPUsresults = list(transcribe.map(audio_urls))

Why it matters

The old way ships YAML.
Velar ships models.

This is the actual difference between setting up GPU infra yourself and using Velar.

Without Velar

Traditional GPU deployment

AWS / GCP / bare metal

# Dockerfile (~40 lines)FROM nvidia/cuda:12.1.0-base-ubuntu22.04RUN apt-get update && apt-get install -y python3 ...COPY requirements.txt .RUN pip install -r requirements.txt # kubernetes.yaml (~80 lines)apiVersion: apps/v1kind: Deploymentspec: template: spec: nodeSelector: cloud.google.com/gke-accelerator: ... # + ECR push · IAM roles · VPC config · autoscaling ...

Setup time: 2–6 hours·Config lines: 200+

With Velar

Pure Python deployment

velar-sdk — pip install velar

import velar app = velar.App("inference")image = velar.Image.from_registry("pytorch/pytorch:2.4.0-cuda12.4") .pip_install("transformers") @app.cls(gpu="A100", image=image) # model loads ONCEclass TextModel: def __enter__(self): from transformers import pipeline self.pipe = pipeline("text-generation", model="meta-llama/Llama-3.1-8B") ✓ app.deploy() # that's it

Setup time: < 60 seconds·Config lines: 0

Features

Built for ML engineers,
not DevOps teams

Every design decision in Velar optimizes for shipping models fast, not managing infrastructure.

@app.cls: Warm models, instant calls

0s model reload after first call

Decorate your class with @app.cls. __enter__ runs once when the container starts — your model stays in GPU memory for every subsequent call. No cold model loads. No wasted compute.

.map(): Parallel GPU fleet

50 GPUs in parallel, out of the box

Call .map() on any function with a list of inputs. Velar provisions one GPU per item and runs them all in parallel. Embed 50k documents, transcribe 1000 audio files, or generate 10k images — same code, any scale.

Volume.from_name(): Cache once, run forever

16GB LLaMA downloaded exactly once

Declare a Volume and mount it on any function. HuggingFace models, datasets, checkpoints — download once to persistent storage, reused on every run. Never wait 5 minutes for a model download again.

Secret.from_name(): Zero credential leaks

Secrets never touch your source code

Store HF_TOKEN, API keys, or any credential in Velar's encrypted vault. Reference by name in your function — the value is injected at runtime and never appears in your code, image, or git history.

Use cases

What people build
with Velar

faster-whisper · .map()

Audio transcription

Batch Whisper jobs on L4. Process 8 hours of audio in 4 minutes.

Learn more

sentence-transformers · .map()

Embedding pipelines

Run sentence-transformers across millions of documents in parallel.

Learn more

diffusers · A100

Image generation

Stable Diffusion batch jobs for product mockups and synthetic datasets.

Learn more

peft · Volume.from_name()

Model fine-tuning

LoRA fine-tuning on your custom dataset. Checkpoints to persistent Volume.

Learn more

@app.cls · persistent endpoint

Custom inference APIs

Deploy your fine-tuned model as a persistent endpoint with a stable URL.

Learn more

@app.function · .map()

Research experiments

Run ablations in parallel with .map(). No cluster setup.

Learn more

Early days

Be one of the first to build on Velar

Velar is new — and that works in your favor. You get founder-level support, direct access to the team, and per-second pricing built to win you over, not to maximize margin.

Start for free

Pricing

GPU by the second.

No minimums. No reserved capacity. The price shown is what you pay.

L40S 48GB

48 GB VRAM

$1.08

per hour

$0.000299 / sec

H100 PCIe 80GB

80 GB VRAM

$2.99

per hour

$0.000830 / sec

H200 141GB

141 GB VRAM

$4.49

per hour

$0.001247 / sec

See full pricing breakdownAll GPUs billed per-second · no minimums

FAQ

Common questions

Can I use any base Docker image?

Yes — any public registry image works. Velar generates the Dockerfile from your function automatically, so most users never write one. If you need a custom base, pass it to Image.from_registry() and Velar layers your deps on top.

What if my job runs longer than expected?

Credits are reserved upfront based on a configurable timeout. When the job finishes or you cancel it, the unused portion is refunded immediately. You only pay for the seconds the GPU was actually running.

What happens after my $10 free credits?

Add a payment method and continue at the same per-second rates. No minimum spend, no reserved capacity, no contracts. You can set a hard monthly spend limit in the dashboard.

How does the image cache work?

Velar uses a content hash of your Dockerfile and handler code. If nothing changed since the last deploy, the build step is skipped entirely — typically under 15 seconds for a warm redeploy.

Ship your
first model
in 60 seconds.

$10 in GPU credits when you sign up. Cancel at any time.

Deploy your first model Read docs

Your Pythonfunction.Any GPU.In minutes.

The old way ships YAML.Velar ships models.

Built for ML engineers,not DevOps teams

What people buildwith Velar

Be one of the first to build on Velar

GPU by the second.

Common questions

Ship yourfirst modelin 60 seconds.

Your Python
function.
Any GPU.
In minutes.

The old way ships YAML.
Velar ships models.

Built for ML engineers,
not DevOps teams

What people build
with Velar

Ship your
first model
in 60 seconds.