Create Deployment
Create a dedicated Fireworks deployment by POSTing a baseModel and a validated deploymentShape, with example curl request and shape lookup guidance.
Creating a deployment with a deployment shape
Deployment shapes are pre-configured templates optimized for speed, cost, or efficiency. To create a deployment with a specific shape, pass thedeploymentShape field in the request body along with baseModel.
Use the Match Deployment Shape Versions endpoint to find available shapes for your model: it takes a deployment create request and returns the validated shape versions compatible with that model, ready to pass as deploymentShape.
activeModelVersion or targetModelVersion — the shape provides the necessary configuration.Always pass deploymentShape
Shapes are validated, so the hardware, precision, and serving configuration are known to work together. Omitting deploymentShape — whether or not you set acceleratorType, acceleratorCount, or precision — creates the deployment without a shape. Deployments without a shape are the most common cause of failed deployment creations, and shapeless creation will soon require an explicit opt-in: always pass deploymentShape. See What is a deployment shape? for how to find shapes for your model.
deploymentShape also accepts the special value default: the server picks a validated shape compatible with the model and your request, and the deployment is created from it. If every compatible shape conflicts with fields you set, or no shape is compatible with the model, the request fails with an error naming the conflicting fields and the compatible shapes — the pick never silently overrides your fields or falls back to a shapeless create.
Each model is validated only on the accelerator type, GPU count, and precision combinations covered by its shapes — many models support just one. Check Match Deployment Shape Versions before overriding acceleratorType or acceleratorCount: an unsupported combination typically fails at creation with a generic Internal error occurred message that does not name the hardware mismatch.
Creating a deployment without a shape (advanced users only)
Pass the query parameteracceptShapelessRisk=true to explicitly create a deployment without a shape. It cannot be combined with deploymentShape — a deployment is either shaped or explicitly shapeless. Shape-specific fields like acceleratorType, acceleratorCount, and precision can still be set:
default) now unless you deliberately need this path. If you need a configuration no shape covers, contact us and we’ll help you find or add one.Authorizations
Query Parameters
Body
The properties of the deployment being created.
Annotations to identify deployment properties. Key/value pairs may be used by external tools or other services. Customer-managed keys supplied during deployment creation must begin with "custom/"; bare keys return HTTP 403 (PERMISSION_DENIED). See Deployment Tags for tag management and migration guidance. Responses to regular account users include only customer-owned tags with keys under the "custom/" prefix.
The type of accelerator to use.
ACCELERATOR_TYPE_UNSPECIFIED, NVIDIA_A100_80GB, NVIDIA_H100_80GB, AMD_MI300X_192GB, NVIDIA_A10G_24GB, NVIDIA_A100_40GB, NVIDIA_L4_24GB, NVIDIA_H200_141GB, NVIDIA_B200_180GB, AMD_MI325X_256GB, AMD_MI350X_288GB, NVIDIA_B300_288GB, NVIDIA_GB200, NVIDIA_GB300 The maximum number of concurrent (in-flight) requests a single replica will accept before shedding load. Requests that arrive while a replica is already at this limit are rejected early with HTTP 429 instead of queueing — a per-replica admission gate for controlling tail latency. When unset (0), the platform default is used.
The draft model name for speculative decoding. e.g. accounts/fireworks/models/my-draft-model If empty, speculative decoding using a draft model is disabled. Default is the base model's default_draft_model. Set CreateDeploymentRequest.disable_speculative_decoding to false to disable this behavior.
Controls hot-load reload semantics for this deployment. ASYNC (default when enable_hot_load is set) pauses the generator/prefiller mid-flight and skips draining in-flight requests during hot load. SYNC drains in-flight requests before applying the new model.
HOT_LOAD_TRANSITION_TYPE_UNSPECIFIED, ASYNC, SYNC Response
A successful response.
Annotations to identify deployment properties. Key/value pairs may be used by external tools or other services. Customer-managed keys supplied during deployment creation must begin with "custom/"; bare keys return HTTP 403 (PERMISSION_DENIED). See Deployment Tags for tag management and migration guidance. Responses to regular account users include only customer-owned tags with keys under the "custom/" prefix.
The type of accelerator to use.
ACCELERATOR_TYPE_UNSPECIFIED, NVIDIA_A100_80GB, NVIDIA_H100_80GB, AMD_MI300X_192GB, NVIDIA_A10G_24GB, NVIDIA_A100_40GB, NVIDIA_L4_24GB, NVIDIA_H200_141GB, NVIDIA_B200_180GB, AMD_MI325X_256GB, AMD_MI350X_288GB, NVIDIA_B300_288GB, NVIDIA_GB200, NVIDIA_GB300 The maximum number of concurrent (in-flight) requests a single replica will accept before shedding load. Requests that arrive while a replica is already at this limit are rejected early with HTTP 429 instead of queueing — a per-replica admission gate for controlling tail latency. When unset (0), the platform default is used.
The draft model name for speculative decoding. e.g. accounts/fireworks/models/my-draft-model If empty, speculative decoding using a draft model is disabled. Default is the base model's default_draft_model. Set CreateDeploymentRequest.disable_speculative_decoding to false to disable this behavior.
The handle for calling a direct route. The meaning of the handle depends on the direct route type of the deployment: INTERNET -> The host name for accessing the deployment GCP_PRIVATE_SERVICE_CONNECT -> The service attachment name used to create the PSC endpoint. AWS_PRIVATELINK -> The service name used to create the VPC endpoint.
The geographic region where the deployment is presently located. This region may change
over time, but within the placement constraint.
REGION_UNSPECIFIED, US_IOWA_1, US_VIRGINIA_1, US_VIRGINIA_2, US_ILLINOIS_1, AP_TOKYO_1, US_ARIZONA_1, US_TEXAS_1, US_ILLINOIS_2, EU_FRANKFURT_1, US_TEXAS_2, EU_ICELAND_1, EU_ICELAND_2, US_WASHINGTON_1, US_WASHINGTON_2, US_WASHINGTON_3, AP_TOKYO_2, US_CALIFORNIA_1, US_UTAH_1, US_ARIZONA_3, US_GEORGIA_1, US_GEORGIA_2, US_WASHINGTON_4, US_GEORGIA_3, NA_BRITISHCOLUMBIA_1, US_GEORGIA_4, US_OHIO_1, US_NEWYORK_1, EU_NETHERLANDS_1, US_WASHINGTON_5, US_MINNESOTA_1, US_CALIFORNIA_2, NA_BRITISHCOLUMBIA_2, AP_MALAYSIA_2, US_OREGON_1, NA_BRITISHCOLUMBIA_3, AP_NEWSOUTHWALES_1, AP_TAIWAN_1 Controls hot-load reload semantics for this deployment. ASYNC (default when enable_hot_load is set) pauses the generator/prefiller mid-flight and skips draining in-flight requests during hot load. SYNC drains in-flight requests before applying the new model.
HOT_LOAD_TRANSITION_TYPE_UNSPECIFIED, ASYNC, SYNC