GPU runtime for your Python — not Replicate's model registry
Replicate is a marketplace for pre-built models. Velar is a serverless runtime for your Python code. No Cog, no model packaging, no registry lock-in — just decorate a function and deploy.
Why developers switch from Replicate to Velar
Replicate works great if you want to call someone else's model via API. If you want to run your own Python — custom models, private fine-tunes, full training pipelines — you need a different tool.
No Cog, no restrictions
Replicate forces you into their Cog build system. Velar lets you bring any Docker image or pip install list. Your workflow, your dependencies — no repackaging.
Your code on any GPU
Replicate is a marketplace for pre-built models. Velar is a runtime for YOUR Python. Decorate a function with @app.function(gpu='A100') and it runs on GPU — no model registry involved.
No vendor lock-in
Replicate owns the model lifecycle. With Velar your code lives in your repo. Switching GPUs, swapping models, or migrating off Velar later is just a Python refactor.
Velar vs Replicate
| Feature | Velar | Replicate |
|---|---|---|
| Run arbitrary Python code on GPUReplicate requires packaging models in Cog | ||
| No proprietary build tool requiredReplicate mandates Cog for all deployments | ||
| Per-second billing | ||
| Scale to zero | ||
| Python-native SDKReplicate's Python client calls hosted models, not your own code | ||
| Works with any ML frameworkReplicate focuses on pre-built public models | ||
| Custom training & fine-tuning jobs | ||
| Persistent endpointsReplicate has no always-on custom endpoint offering | ||
| Free starter credits |
Your Python. Any GPU. No Cog.
Replicate requires you to wrap your model in Cog, define an input/output schema, and push to their registry. With Velar, you decorate a Python function and call app.deploy(). That is the entire workflow.
- Any pip-installable dependency works
- Private model weights — no public registry required
- Fan out with .map() across many GPUs in parallel
- Fine-tuning and training jobs, not just inference
import velar
app = velar.App("my-pipeline")
# Any image, any dependencies — no Cog required
image = velar.Image.from_registry(
"pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime"
).pip_install("diffusers", "transformers", "accelerate")
@app.function(gpu="L40S", image=image)
def run(prompt: str) -> bytes:
from diffusers import StableDiffusionXLPipeline
import torch, io
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
).to("cuda")
buf = io.BytesIO()
pipe(prompt).images[0].save(buf, "PNG")
return buf.getvalue()
# Fan out across many GPUs in parallel
app.deploy()
prompts = load_prompts() # your list
results = [run.remote(p) for p in prompts]Frequently asked questions
Do I need to package my model as a Cog container?
No. Velar has no equivalent of Cog. You specify a base Docker image and pip packages in Python code — no separate build tool, no cog.yaml, no schema.yaml.
Can I run models not available on Replicate's platform?
Yes. Velar runs any Python code on any GPU. If you can install it with pip or conda, you can run it on Velar — internal models, fine-tuned checkpoints, and private weights included.
How does Velar pricing compare to Replicate?
Velar bills transparent per-second rates — cheaper than Modal on every GPU. L40S runs from $1.08/hr, A100 80GB from $1.74/hr, and billing stops the instant your function finishes. Replicate bills per prediction with per-model pricing that varies widely.
Can I run fine-tuning and training jobs, not just inference?
Yes. Unlike Replicate which focuses on inference, Velar supports long-running GPU jobs up to 24 hours — fine-tuning, batch processing, embedding generation, anything.
Does Velar have a model marketplace like Replicate?
No, and that is by design. Velar is a GPU runtime for your own code. If you want public models, Replicate is a good fit. If you want to run your own Python workloads, Velar is built for you.
Run your Python on GPU — not someone else's model
Velar Cloud is paused. Join the waitlist for the relaunch. Deploy your first workload in under 5 minutes.