Generative AI in CPU-Only Regions
Use CPU-only access and routing profiles to run on-demand inference on models hosted in other OCI regions.
Use Generative AI from your selected region even when the model you want is hosted in another region. With CPU-only access, on-demand inference runs on models hosted in other regions. Create a routing profile for the model and select one or more GPU regions where it is available. Then select that profile in the Playground or reference it in an API request. Generative AI routes the request to a selected region to run the model.
For example, in Singapore (ap-singapore-1), you can create a routing profile for openai.gpt-oss-120b with Frankfurt (eu-frankfurt-1) as its target region. Select the profile in Singapore's Playground to chat with the model served from Frankfurt. You can continue working in Singapore without switching the Console to Frankfurt.
Model inference through routing profiles is available in on-demand mode only. Each profile specifies one supported model and one or more target regions in the same realm where that model is available.
Routing profiles are also available in regions where OCI Generative AI hosts models. For example, if an on-demand model is available in US Midwest (Chicago) but not US East (Ashburn), create a routing profile in Ashburn and select Chicago as its target region. Then select the profile in Ashburn's Playground or reference it in a request to the Ashburn API endpoint. Generative AI routes the inference request to Chicago.
The Chat API, OpenAI-compatible Chat Completions API, and Responses API support routing profiles for models that support the selected API. For Responses API model and region availability, see Models and Regions for Enterprise AI Agents. Generative AI API endpoints in CPU-only regions support IPv6.
For a walkthrough of creating a routing profile and using it in Chat Completions and Responses API requests, see the tutorial Route Model Requests Across OCI Regions.
For model availability in target regions, see Models by Region. The feature tables below list the functionality available through CPU-only access in the regions listed on this page, including UAE Central (Abu Dhabi).
Access a Model in Another Region
Create the routing profile in the region where you want to use Generative AI. The profile selects the model and its serving regions; selecting the profile in the Playground lets you use that model from your current region.
- In the Console, select your CPU-only region, such as Singapore (Singapore) (
ap-singapore-1). - Open Analytics & AI, select Generative AI, and then select Routing profiles.
- Select Create routing profile. Enter a name, such as
singapore-frankfurt, and select a compartment. - Under Configuration, select the model you want to use from Model select, such as
openai.gpt-oss-120b. Target OCI regions lists the GPU regions where that model is available. Select one or more of those regions, such as Frankfurt (eu-frankfurt-1), and select Create. For complete instructions, see Creating a Routing Profile. - Open the profile's details page and confirm that its status is Active. Review Model selection and Target OCI regions to confirm the model and serving regions.
- While still in your CPU-only region, open Playground, select Chat, and select the profile's compartment. In the Model menu, select your routing profile, such as
singapore-frankfurt. - Enter a prompt and select Submit. Generative AI runs the model specified in the profile in one of its selected target regions. In the Playground, review Routed model and Serving region to see which model and region served the request.
You can use the same profile in API requests sent to your region's endpoint. For an OpenAI-compatible Chat Completions or Responses request, set model to the profile's full OCID. Select a model that supports the API you plan to call. See Route Model Requests Across OCI Regions for Python examples and setup instructions for both APIs.
Commercial Regions (OC1) with CPU-Only Access
The following table lists the OC1 regions that support CPU-only access for on-demand inference in Generative AI. Region names, identifiers, and keys follow the OCI regions table.
UAE Central (Abu Dhabi) also hosts models for dedicated AI clusters. On-demand inference from Abu Dhabi uses a routing profile to access a supported model in another region. For locally hosted models and their dedicated availability, see Models by Region: Middle East.
| Region Name | Location | Region Identifier | Region Key |
|---|---|---|---|
| Australia East (Sydney) | Sydney |
ap-sydney-1
|
SYD
|
| Australia Southeast (Melbourne) | Melbourne |
ap-melbourne-1
|
MEL
|
| Brazil Southeast (Vinhedo) | Vinhedo |
sa-vinhedo-1
|
VCP
|
| Canada Southeast (Montreal) | Montreal |
ca-montreal-1
|
YUL
|
| Canada Southeast (Toronto) | Toronto |
ca-toronto-1
|
YYZ
|
| Chile Central (Santiago) | Santiago |
sa-santiago-1
|
SCL
|
| Chile West (Valparaiso) | Valparaiso |
sa-valparaiso-1
|
VAP
|
| Colombia Central (Bogota) | Bogota |
sa-bogota-1
|
BOG
|
| France Central (Paris) | Paris |
eu-paris-1
|
CDG
|
| India West (Mumbai) | Mumbai |
ap-mumbai-1
|
BOM
|
| Indonesia North (Batam) | Batam |
ap-batam-1
|
HSG
|
| Italy Northwest (Milan) | Milan |
eu-milan-1
|
LIN
|
| Italy North (Turin) | Turin |
eu-turin-1
|
NRQ
|
| Japan East (Tokyo) | Tokyo |
ap-tokyo-1
|
NRT
|
| Mexico Central (Queretaro) | Queretaro |
mx-queretaro-1
|
QRO
|
| Mexico Northeast (Monterrey) | Monterrey |
mx-monterrey-1
|
MTY
|
| Morocco West (Casablanca) | Casablanca |
af-casablanca-1
|
LEJ
|
| Netherlands Northwest (Amsterdam) | Amsterdam |
eu-amsterdam-1
|
AMS
|
| Singapore (Singapore) | Singapore |
ap-singapore-1
|
SIN
|
| Singapore West (Singapore) | Singapore |
ap-singapore-2
|
XSP
|
| South Africa Central (Johannesburg) | Johannesburg |
af-johannesburg-1
|
JNB
|
| South Korea Central (Seoul) | Seoul |
ap-seoul-1
|
ICN
|
| South Korea North (Chuncheon) | Chuncheon |
ap-chuncheon-1
|
YNY
|
| Spain Central (Madrid) | Madrid |
eu-madrid-1
|
MAD
|
| Spain Central (Madrid 3) | Madrid |
eu-madrid-3
|
ORF
|
| Sweden Central (Stockholm) | Stockholm |
eu-stockholm-1
|
ARN
|
| Switzerland North (Zurich) | Zurich |
eu-zurich-1
|
ZRH
|
| UAE Central (Abu Dhabi) | Abu Dhabi |
me-abudhabi-1
|
AUH
|
| UK West (Newport) | Newport |
uk-cardiff-1
|
CWL
|
| US West (San Jose) | San Jose |
us-sanjose-1
|
SJC
|
Features for Enterprise AI Models
| Feature | Availability |
|---|---|
| Regional routing with routing profiles | Available for supported on-demand models. Calls are routed to GPU regions. |
| Chat API | Available through a routing profile for models that support the Chat API. |
| Google Gemini inference | Available through a routing profile for supported model and region combinations. |
| xAI Grok inference | Available through a routing profile for supported model and region combinations. |
| Meta and Cohere inference | Available through a routing profile. Embedding requests aren't supported. |
| OpenAI gpt-oss-120b inference | Available through a routing profile for supported model and region combinations. |
| Model retirement | Available. |
| Dedicated AI clusters and model endpoints | Not available. |
| Model import and imported model inference | Not available. |
| Embedding API | Not available. |
| Private endpoints for Google Gemini and xAI Grok models | Available. |
| Private endpoints for hosted or imported models | Not available. |
Features for Enterprise AI Agents
| Feature | Availability |
|---|---|
| Sandboxes | Available. |
| Hosted deployments | Available. |
| Projects | Available with short-term memory. Long-term memory isn't supported. |
| Short-term memory | Available. |
| Semantic stores and vector stores | Not available. |
| File search | Not available. |
| OpenAI-compatible Chat Completions API | Available through a routing profile for models that support Chat Completions. |
| Responses API | Available through a routing profile for models that support the Responses API. Tools and other Enterprise AI Agents features are subject to the availability listed in this table. |
| Agent gateway for hosting MCP tools | Not available. |
| Client-side MCP tools | Not available. |
| Long-term memory | Not available. |
| Natural language to SQL (NL2SQL) | Not available. |
Features for Enterprise AI Governance
| Feature | Availability |
|---|---|
| API keys | Available. |
| Guardrails | Not available. |