Skip to content

Canada (Central) Region Model Support

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. SPDX-License-Identifier: MIT-0

The accelerator’s region-aware model handling covers Canada. When the stack is deployed in a ca-* region, the UpdateConfiguration custom resource (src/lambda/update_configuration/index.py) resolves the stack region to its own region type, ca, and applies the two deploy-time steps it already applies for US and EU stacks: it limits the model picklists to models that can be called from ca-central-1, and it swaps the shipped defaults that cannot. us-gov-* regions keep resolving to the US type.

The handling has been validated in ca-central-1 (live checks, 2026-10). ca-west-1 receives the same handling but has not been validated. The EU counterpart, with its full US-to-EU mapping table, is EU Region Model Support and Mapping.

Every model / model_id picklist in the configuration schema is filtered by id prefix when the stack is created or updated:

IdsIn a Canada stackWhy
ca.*keptCanada geo inference profile
us.*, global.*, unprefixed ids, LambdaHookkept, except the deny list belowcallable from ca-central-1
eu.*droppednot callable from ca-central-1
openai.gpt-5.4, openai.gpt-5.5, openai.gpt-5.6-sol / -terra / -lunadroppedbedrock-mantle ids with no Canada availability; the client would route them to us-east-1

The deny list (CA_UNAVAILABLE_MODELS) holds the shipped ids that have no inference profile and no in-region hosting in ca-central-1 (verified 2026-10-11):

  • us.amazon.nova-lite-v1:0 — swapped to ca.amazon.nova-lite-v1:0, see below
  • us.amazon.nova-pro-v1:0 and global.amazon.nova-pro-v1:0
  • us.cohere.embed-v4:0 — swapped to global.cohere.embed-v4:0, see below
  • us.meta.llama4-maverick-17b-instruct-v1:0, us.meta.llama4-scout-17b-instruct-v1:0
  • qwen.qwen3-vl-235b-a22b, google.gemma-3-27b-it, nvidia.nemotron-nano-12b-v2

us.xai.grok-4.6 and us.openai.gpt-6-astra have profiles in ca-central-1 and stay in the picklists, as do the :priority and :flex service-tier variants of us.amazon.nova-2-lite-v1:0 and global.amazon.nova-2-lite-v1:0, which Bedrock accepts from ca-central-1 (verified 2026-10-12).

Three shipped defaults name models that cannot be called from ca-central-1. On stack create and update, the custom resource rewrites them in the ConfigSchema, the shipped default configuration and the managed configurations through a one-way Canada table (CA_MODEL_MAPPINGS). Configuration profiles you saved yourself are not rewritten, the same as in EU regions.

SettingShipped defaultIn a Canada stackWhy
Confidence assessment (extraction.confidence.model)us.amazon.nova-lite-v1:0ca.amazon.nova-lite-v1:0Same model through its Canada geo profile; processes in Canada
Discovery, with and without ground truth (model_id)us.amazon.nova-pro-v1:0us.amazon.nova-2-lite-v1:0No Nova Pro profile in ca-central-1. Nova 2 Lite is the catalog’s Nova default for OCR, classification and evaluation, and is multimodal and cheaper
Multi-document discovery embedding (multi_document.embedding_model_id)us.cohere.embed-v4:0global.cohere.embed-v4:0Only the global profile exists in Canada. The swap is ca-central-1-only in effect: this profile does not exist in ca-west-1 either, so a ca-west-1 stack moves from one id it cannot call to another

The two text-generation ids swapped in above are in the Canada picklists. multi_document.embedding_model_id has no picklist — it is a free-text field with a default: and no enum — so its value is only ever set by this swap or by hand.

ca.amazon.nova-lite-v1:0 is new to the catalog with this support, and it is the only ca. profile AWS offers. It is in every model dropdown where its EU twin is, in the KnowledgeBaseModelId AllowedValues, priced with Canada (Central) rates ($0.064 input, $0.256 output and $0.016 cache read per 1M tokens; cache writes are billed at the input rate, like the other Nova rows), and prompt caching is enabled for it (CACHEPOINT_SUPPORTED_MODELS). The ca. prefix is recognised by the shared region-prefix list, so the client resolves it to the same base model as the us. and eu. Nova Lite rows.

The stack calls the ca-central-1 Bedrock endpoint for everything, and all data at rest stays in ca-central-1. Cross-Region inference, however, may process a request outside Canada; which Regions depends on the profile prefix:

ProfileRequests processed in
ca.*ca-central-1, ca-west-1
us.*US Regions
global.*any commercial AWS Region
BDA na.data-automation-v1ca-central-1, ca-west-1, us-east-1, us-east-2, us-west-2

ca.amazon.nova-lite-v1:0 is therefore the only model in the catalog whose requests stay in Canada. See Cross-Region inference and Cross Region support required for Bedrock Data Automation in the Amazon Bedrock User Guide.

Canada stacks use the na.data-automation-v1 cross-Region profile — AWS names the Canada (Central) profile na., and ca.data-automation-v1 is not a valid profile — so BDA mode (use_bda, which calls the asynchronous API) works.

The BDA OCR backend (ocr.backend: bda) does not. It needs a stack-scoped SYNC document project, and BDA’s synchronous API rejects that project in ca-central-1 with a ValidationException. The BDAOCRProjectFunction custom resource treats the refusal as “unsupported here”: it returns success with an empty project ARN and logs one ERROR line (Bedrock Data Automation rejected the stack-scoped OCR project as unsupported in this region ...; returning empty ProjectArn). That line is expected, and the stack deploys. The OCR service raises a clear error only if ocr.backend: bda is selected; use the Textract (default) or Bedrock OCR backend instead (see When to choose BDA vs Textract for OCR). An asynchronous OCR path is a possible follow-up.

KnowledgeBaseModelId is a stack parameter, so the configuration step does not swap it. Its default, us.amazon.nova-pro-v1:0, has no inference profile in ca-central-1, so the root template swaps it instead: a stack in a Canada region that leaves the parameter at its default gives the Knowledge Base ca.amazon.nova-lite-v1:0, which keeps Knowledge Base answers in Canada (the stack parameter itself still shows the default). An explicit choice is passed through unchanged, so it must be one ca-central-1 can call, such as us.amazon.nova-2-lite-v1:0, us.anthropic.claude-sonnet-4-5-20250929-v1:0 or global.anthropic.claude-sonnet-4-5-20250929-v1:0. The default Knowledge Base embedding model (KnowledgeBaseEmbeddingModel), amazon.titan-embed-text-v2:0, is supported in ca-central-1; confirm a Cohere alternative is offered in the region before selecting it. See Document Knowledge Base Query.

  • Capacity planner. calculate_capacity has no Service Quotas keys for ca. (or eu.) profiles, so capacity estimates do not cover ca.amazon.nova-lite-v1:0. This is a pre-existing gap shared with EU; see Capacity Planning.
  • GPT-5.x ids typed into a custom configuration. The picklists hide them, but an openai.gpt-5.* id written into a configuration by hand still routes to us-east-1 through bedrock-mantle, with only a warning logged (same as EU). Avoid them in Canada.
  • The Canada swaps are one-way. A configuration that holds a ca. id keeps it when carried into a US or EU stack, and an eu. id stored in a Canada stack is not rewritten. The picklists hide both, but a stored value stays until someone edits it.
  • Embedding calls are not costed. Multi-document discovery’s embedding model (global.cohere.embed-v4:0 in Canada, eu.cohere.embed-v4:0 in EU Regions, us.cohere.embed-v4:0 elsewhere) emits no metering entry — generate_embedding returns a bare vector — and none of these ids has a row in config_library/pricing.yaml, so embedding usage does not appear in cost reporting. Pre-existing and not specific to Canada. The coverage gates record this per id rather than as a blanket rule for embeddings: _COVERAGE_EXEMPT in lib/idp_common_pkg/tests/unit/reporting/test_pricing_lookup.py and EMBEDDING_MODELS in scripts/tests/test_model_surface_consistency.py name the Knowledge Base embedding models, and NON_SELECTABLE_DEFAULTS names us.cohere.embed-v4:0. global.cohere.embed-v4:0 and eu.cohere.embed-v4:0 are in none of them. They reach a stack two ways that those gates do not walk: the swap table, and the Region-aware code fallbacks in idp_common/regional_models.py. The fallbacks are held to the swap table instead, by src/lambda/update_configuration/test_eu_region.py. A shipped preset or a config/models.py default naming either id would fail the coverage gates with no entry to point at.
  • ca-west-1 is handled, but the model lists are derived from ca-central-1. A ca-west-1 stack resolves to the same ca region type and gets the same picklists, swaps and Knowledge Base default, which is validated for ca-central-1 only. Checked live on 2026-10-13: of the ids the Canada picklists offered then, four could not be called from ca-west-1 — global.xai.grok-4.6, us.anthropic.claude-haiku-4-5-20251001-v1:0, us.anthropic.claude-opus-4-5-20251101-v1:0 and us.anthropic.claude-sonnet-4-5-20250929-v1:0 — and global.cohere.embed-v4:0, the multi-document discovery embedding default, has no ca-west-1 profile. ca.amazon.nova-lite-v1:0, which the Knowledge Base condition selects, is callable there. Verify with the command in Verifying model availability against your own Region before relying on a model.
  • Claude Sonnet 5.5 is offered in Canada and its ca-west-1 status is open. us.anthropic.claude-sonnet-5-5 and global.anthropic.claude-sonnet-5-5 are both ACTIVE in ca-central-1 and the bare anthropic.claude-sonnet-5-5 is hosted in-Region there (checked 2026-10-14), which is why the model carries no CA_UNAVAILABLE_MODELS entry. Neither that nor its two Service Quotas codes (L-94A31E46, L-31AB82D0) has been resolved from ca-west-1. The quota codes matter more than the usual picklist question there: CalculateCapacity re-raises any non-throttling error from get_service_quota, so a code that does not resolve fails the whole Capacity Planning response rather than one row. Run the availability command below plus aws service-quotas get-service-quota --service-code bedrock --quota-code L-94A31E46 from a ca-west-1 stack before relying on either.

ca-central-1 is one of the four Regions each release is published to, so it has a Launch Stack button in the Deployment Guide and a plain idp-cli deploy --region ca-central-1 resolves the published template without your having to name one. ⚠️ 0.6.11 and every release before it have no ca-central-1 artifacts — Canada joined the published set in 0.6.12, so both routes work only from that release on. To deploy an earlier release, or unreleased code, supply the template yourself in one of two forms:

Terminal window
# Build, publish to a bucket in your account and deploy in one step
idp-cli deploy --stack-name my-idp --region ca-central-1 --from-code . \
--admin-email your.email@example.com --wait
# Or publish first, then deploy the published template
idp-cli publish --source-dir . --region ca-central-1
idp-cli deploy --stack-name my-idp --region ca-central-1 \
--template-url <template URL printed by publish> \
--admin-email your.email@example.com --wait

The general build-from-source procedure, including the prerequisites, is in the Deployment Guide.

The authoritative list of what ca-central-1 can call is the set of active system-defined inference profiles:

Terminal window
aws bedrock list-inference-profiles --region ca-central-1 --type-equals SYSTEM_DEFINED \
--query "inferenceProfileSummaries[?status=='ACTIVE'].inferenceProfileId" --output text \
| tr '\t' '\n' | sort

A prefixed id that is not in this list cannot be called from a Canada stack. The deny list above was built from this command (plus aws bedrock list-foundation-models --region ca-central-1 for the unprefixed ids) on 2026-10-11; run them again before relying on a model this page does not mention.