Models

A model ID is a lowercase, stable string derived from the upstream model name — family, size and variant — for example qwen3-30b-a3b. Pass it in the model field of any request:

• IDs do not encode quantization or runtime. Those are deployment details, documented per deployment and pinned on dedicated nodes.
• GET /v1/models is the source of truth for what your key can call. The catalog below lists what can be deployed.
• Calling a model that exists but is not enabled for your key returns 403 model_not_in_plan; an unknown ID returns 404 model_not_found.

Shared means served on the multi-tenant Managed AI API. Dedicated means deployable on a single-tenant Private AI Node. Context is the model's published maximum; the served context on a deployment can be lower, because the KV cache shares memory with the weights.

Only models marked Available on the shared API appear in /v1/models for shared plans. Private Node and On Request models are deployed on dedicated capacity after a short fit check for your workload.

Open-weight models receive upstream revisions, and quantized builds can be regenerated. Both can change outputs, so production applications need control over when that happens.

• Dedicated nodes: the deployed build — upstream revision, quantization and runtime version — is pinned. It changes only when you agree to an update, ideally after you have re-run your evaluation set against the new build. Aliases (for example support-assistant pointing to a pinned build) can be configured on request so you can switch builds without changing application code.
• Shared API: a model ID refers to a maintained deployment of that model. Changes that can affect outputs are communicated to customers before they take effect. If you need a frozen build, use a dedicated node.

Quantization stores model weights at lower numerical precision — for example 4-bit or 8-bit instead of 16-bit. It reduces the memory a model needs, which determines what fits on a node with 96 GB of unified memory and how much room is left for context.

• 8-bit is generally close to full precision in output quality and needs roughly half the memory of 16-bit.
• 4-bit makes larger models (such as 70B-class) fit on a single node. Quality impact depends on the model and task, so evaluate on your own data.
• Native formats such as MXFP4 for gpt-oss are the format the model is distributed in, not a later compression step.

The quantization of each deployment is documented and, on dedicated nodes, chosen with you. Approximate weight footprints by memory class:

Weight footprints exclude the KV cache, which grows with context length and concurrent requests.