Bedrock models that stay in South Korea

Inference must execute inside South Korea. Of the 28 models the matrix shows as available here, 5 are served without the request leaving, and 23 are not.

Regions counted: Asia Pacific (Seoul) (ap-northeast-2).

What catches people here

Availability in a South Korea region does not mean the call stays there. Check the route column: a model reachable only through a cross-region profile is processed wherever that profile has capacity.

Stays in region 5

Callable on-demand with the bare model ID. The request is served by the region you addressed, and nothing crosses a border.

Model Call this ID Input /1M
Titan Text Embeddings V2 amazon.titan-embed-text-v2:0
Claude 3 Haiku other route leaves anthropic.claude-3-haiku-20240307-v1:0
Claude 3.5 Sonnet other route leaves anthropic.claude-3-5-sonnet-20240620-v1:0
Marengo Embed 3.0 twelvelabs.marengo-embed-3-0-v1:0
Pegasus v1.2 other route leaves twelvelabs.pegasus-1-2-v1:0

This holds for as long as you address the bare model ID. Swapping it for a global. or cross-region profile ID is a one-word change that moves the inference, and the models tagged other route leaves are the ones where that alternative exists.

Leaves the jurisdiction 7

The matrix shows this model as available here. It is only reachable through a profile that declares destinations outside the jurisdiction, so a share of your traffic is processed there.

Model Processed in
Nova Lite Amazon Taiwan, Japan, South Korea, India, Singapore, Australia, Indonesia
Nova Micro Amazon Taiwan, Japan, South Korea, India, Singapore, Australia, Indonesia
Nova Pro Amazon Taiwan, Japan, South Korea, India, Singapore, Australia, Indonesia
Claude 3 Sonnet Anthropic Japan, South Korea, India, Singapore, Australia
Claude 3.5 Sonnet v2 Anthropic Japan, South Korea, India, Singapore, Australia
Claude Sonnet 4 Anthropic Taiwan, Japan, South Korea, India, Singapore, Australia, Indonesia
Marengo Embed v2.7 TwelveLabs Japan, South Korea, India, Singapore, Australia, Indonesia

Destination not bounded 16

The only route is a global profile. AWS does not publish a closed destination list for it, so no residency claim can be made either way — this is unverifiable, not merely non-compliant.

Model Processed in
Nova 2 Lite Amazon not published
Claude Fable 5 Anthropic not published
Claude Haiku 4.5 Anthropic not published
Claude Opus 4.5 Anthropic not published
Claude Opus 4.6 Anthropic not published
Claude Opus 4.7 Anthropic not published
Claude Opus 4.8 Anthropic not published
Claude Opus 5 Anthropic not published
Claude Sonnet 4.5 Anthropic not published
Claude Sonnet 4.6 Anthropic not published
Claude Sonnet 5 Anthropic not published
Embed v4 Cohere not published
GPT-5.6 Luna OpenAI not published
GPT-5.6 Sol OpenAI not published
GPT-5.6 Terra OpenAI not published
Grok 4.6 xAI not published

Not offered here 88

Not available in any region of this jurisdiction, by any route.

List the 88 models

Jamba 1.5 Large , Jamba 1.5 Mini , Amazon Nova Multimodal Embeddings , Nova 2 Sonic , Nova Canvas , Nova Premier , Nova Reel , Nova Reel , Nova Sonic , Rerank 1.0 , Titan Embeddings G1 - Text , Titan Multimodal Embeddings G1 , Titan Text Embeddings v2 , Claude 3.7 Sonnet , Claude Opus 4.1 , Embed English , Embed Multilingual , Rerank 3.5 , DeepSeek V3.2 , DeepSeek-R1 , DeepSeek-V3.1 , Gemma 3 12B IT , Gemma 3 27B PT , Gemma 3 4B IT , Ray v2 , Llama 3 70B Instruct , Llama 3 8B Instruct , Llama 3.1 70B Instruct , Llama 3.1 8B Instruct , Llama 3.3 70B Instruct , Llama 4 Maverick 17B Instruct , Llama 4 Scout 17B Instruct , MiniMax M2 , MiniMax M2.1 , MiniMax M2.5 , Devstral 2 123B , Magistral Small 2509 , Ministral 14B 3.0 , Ministral 3 8B , Ministral 3B , Mistral 7B Instruct , Mistral Large (24.02) , Mistral Large (24.07) , Mistral Large 3 , Mistral Small (24.02) , Mixtral 8x7B Instruct , Pixtral Large (25.02) , Voxtral Mini 3B 2507 , Voxtral Small 24B 2507 , Kimi K2 Thinking , Kimi K2.5 , Nemotron Nano 3 30B , NVIDIA Nemotron 3 Super 120B A12B , NVIDIA Nemotron Nano 12B v2 VL BF16 , NVIDIA Nemotron Nano 9B v2 , GPT OSS Safeguard 120B , GPT OSS Safeguard 20B , gpt-oss-120b , gpt-oss-20b , Qwen3 235B A22B 2507 , Qwen3 32B (dense) , Qwen3 Coder 480B A35B Instruct , Qwen3 Coder Next , Qwen3 Next 80B A3B , Qwen3 VL 235B A22B , Qwen3-Coder-30B-A3B-Instruct , Stable Diffusion 3.5 Large , Stable Image Conservative Upscale , Stable Image Control Sketch , Stable Image Control Structure , Stable Image Core 1.0 , Stable Image Creative Upscale , Stable Image Erase Object , Stable Image Fast Upscale , Stable Image Inpaint , Stable Image Outpaint , Stable Image Remove Background , Stable Image Search and Recolor , Stable Image Search and Replace , Stable Image Style Guide , Stable Image Style Transfer , Stable Image Ultra 1.0 , Palmyra X4 , Palmyra X5 , Writer Palmyra Vision 7B , GLM 4.7 , GLM 4.7 Flash , GLM 5

Questions

How many Bedrock models keep data in South Korea?

5 of the 28 models available in South Korea are served without the request leaving it. The remaining 23 are reachable only through cross-region inference profiles whose declared destinations sit elsewhere, or through a global profile with no published destination list.

Does availability in South Korea mean the model runs there?

No. A model reachable only through a cross-region inference profile is served by whichever region in that profile's scope has capacity. The API call succeeds either way, so availability and residency have to be checked separately — that is what the route column on this page is for.

What exactly is being checked here?

Two facts from the AWS Bedrock APIs, rebuilt daily: whether the bare model ID is callable on-demand in a region of this jurisdiction, and which regions each inference profile declares as destinations. The mapping from region to country is maintained by hand. This describes where inference executes; it is not legal advice and does not cover logs, retention or where your application tier runs.