Skip to main content
Omnia offers a catalog of chat, vision, and embedding models sourced across multiple clouds. Every model has a namespaced id like Qwen/Qwen3-32B, and you call a model by that id. The underlying hardware and provider are an implementation detail: you interact with one consistent API and one catalog.

​
Listing models

Fetch the models available to your workspace at any time with GET /v1/models. The response is an OpenAI-style list of model objects.
GET /v1/models is part of the OpenAI-compatible inference API (https://gateway.omnia-voice.com/v1), not the management API. It uses the same workspace key as every other call.
The full catalog with live per-token pricing is also on the models page in the dashboard.
The models you can call are the ones Omnia has priced and offers. A model that isn’t in the catalog returns 404. This is deliberate, so you’re never billed for something without a published rate.

​
Model ids

Model ids are namespaced by their family, in the form <namespace>/<name>. The namespace groups related models; it is not a provider or cloud name. Use the id exactly as returned by GET /v1/models.

​
Modalities

Models fall into three billable modalities:

​
Example models

A few of the models available at launch (see the dashboard for the full, live list and pricing):
  • Qwen/Qwen3-32B: general chat
  • meta-llama/Llama-3.3-70B-Instruct: general chat
  • openai/gpt-oss-120b: general chat
  • google/gemma-3-27b-it: general chat
  • Qwen/Qwen2.5-VL-72B-Instruct: vision
  • Qwen/Qwen3-Embedding-8B: embeddings

​
Two ways to run a model

The same catalog model can be served in two modes, and you choose per workload:

Shared (per-token)

The default. Requests run on a pooled, multi-tenant fleet and are billed per token, separately for input and output. Nothing to provision; just call the model by its id.

Dedicated (per-GPU-hour)

Reserve private GPU capacity for a specific model and pay per GPU-hour while it runs, not per token. Best for sustained high volume, guaranteed throughput, or serving a fine-tuned model. See Dedicated endpoints.
Both modes are reached through the same inference API. In shared mode you pass the catalog id (model="Qwen/Qwen3-32B"); in dedicated mode you pass model="dedicated/<routing-key>" to route to your own endpoint.

​
The Models page in the dashboard

Beyond the catalog, the dashboard’s Models page is where your own models live:
  • Your models: every fine-tune you’ve deployed and every served model version, in one place: what exists, and where it’s serving from.
  • Aliases: the model aliases manager: create, repoint, or delete a stable name, set up a canary split, and configure its quality gate in the same dialog: judge, mode, sample floor, rollback threshold.

​
Pricing

Shared models are priced per million tokens, separately for input and output. Pricing is transparent and shown per model in the dashboard, and you pay only for the tokens you actually use. Dedicated capacity is billed per GPU-hour instead of per token. See Billing for how charges are metered and settled.

Dedicated endpoints

Reserve private GPU capacity for a model.

Fine-tuning

Train a model on your own data.