Custom models

Your model. Our GPUs.

Fine-tuned model? Proprietary weights? Bring it to PolarGrid. We host it on dedicated GPUs, keep it warm, and hand you an OpenAI-compatible endpoint.

How it works

Three steps to production.

  1. 01

    Share your model

    Send us your weights via HuggingFace, S3, or direct upload. We support most popular architectures.

  2. 02

    We deploy it to our GPU sites

    Your model gets loaded onto dedicated PolarGrid GPUs and kept warm 24/7. No cold starts, no queues.

  3. 03

    You get an endpoint

    OpenAI-compatible API endpoint. Swap it in wherever you currently call an LLM.

Supported

What we can host.

  • Fine-tuned Llama variants
  • Fine-tuned Qwen variants
  • Custom embedding models
  • Proprietary model weights
  • Instruction-tuned models
  • Domain-specific LLMs

Don't see your architecture listed? Get in touch — we evaluate new architectures on a case-by-case basis.

Managed

No GPU ops. Just inference.

Always warm

Your model stays loaded on GPU 24/7. No cold starts, no spin-up delays on the first request.

Multi-site

Deployed across PolarGrid's network of GPU sites in North America, so latency stays low wherever your users are.

OpenAI-compatible

Every hosted model gets an OpenAI-compatible endpoint. No SDK changes on your end.

Dedicated capacity

Your model runs on reserved GPU capacity — not shared with other tenants at peak time.

Questions

Custom model hosting, answered.

Which model formats can I bring?
Share weights from a HuggingFace repository, an S3 bucket, or a direct upload. We support most popular open architectures, including fine-tuned Llama and Qwen variants and embedding models. If your architecture is unusual, tell us what you have and we will evaluate it.
Do I need to change my application code?
Usually not. Every hosted model is served behind an OpenAI-compatible API, so most teams switch by changing the base URL and model name in the client they already use.
Why does keeping a model warm matter?
Cold starts add seconds to the first request while weights load onto the GPU. For real-time products such as chat and agents, that delay is noticeable. We keep your model loaded on dedicated capacity so the first request is as fast as the hundredth.
Where does my model run?
Your model is deployed to PolarGrid sites and requests are routed to the best available site, which keeps network latency low for a geographically spread audience.
Is my model shared with other customers?
No. Custom models run on reserved GPU capacity for your workload, so other tenants cannot crowd out your traffic at peak times.
How is custom hosting priced?
Pricing depends on model size and the throughput you need. Contact our team with your model and expected traffic and we will scope a plan.

Ready to bring your model?

Pricing is based on model size and throughput requirements. Reach out and we'll scope it out.

Talk to us