さくらのAI Engine
: How to use it, features, and the business problems it solves

What is さくらのAI Engine (Sakura AI Engine)?

さくらのAI Engine (Sakura AI Engine) is a generative AI inference API service from SAKURA internet Inc. You send a request to foundation models such as LLMs running on SAKURA internet's GPU servers, and the inference result comes back through the API. You do not need to rent GPUs or build an environment to run the models yourself.

Its defining feature is that model execution and communication stay entirely within data centers in Japan. The company commits to Zero Data Retention (ZDR): inputs and responses are not stored and are not used for model training or service improvement. This makes it easier to consider for companies whose internal rules forbid sending data to overseas APIs.

There are two offerings.

The Private Edition launched on October 8, 2026. Pricing and models on this page are based on the official website as of October 21, 2026.

How to use it

1. Create a SAKURA internet member ID The member ID is free. Registration requires phone number verification. If you already have one, skip to the next step.
2. Create a "Sakura Cloud" project Log in to the control panel and create a Sakura Cloud project. Sakura Cloud requires a registered credit card (not needed if you have already registered one for another SAKURA internet service).
3. Sign up for a Sakura AI Engine plan In the control panel, choose either the "Foundation Model Free Plan" or the "Pay-as-you-go Plan".
4. Issue an API account token Issue a token in the control panel and put it into your app or tool settings. Because the API is OpenAI-compatible, you can use the official OpenAI libraries as-is.
5. Try models in the Playground Send the same question to several models in your browser and compare response speed and quality. It helps you pick a model before writing code.

The control panel also shows the current month's charges and, for the last 30 days, tokens and requests by model.

Features

• Text generation (Chat Completions) LLM text generation for chat, summarization, writing, and more.
• Audio transcription Converts speech to text with whisper-large-v3-turbo.
• Text vectorization (Embeddings) Converts text into vectors with models such as multilingual-e5-large for search and classification.
• Text-to-Speech Reads text aloud in eight VOICEVOX voices, including Zundamon, Shikoku Metan, and Kasukabe Tsumugi. Usage terms follow each VOICEVOX character's license.
• Document management and search (RAG) Register internal documents and have the LLM answer based on retrieved content, with the mechanism provided on the API side.
• Playground Compare and try multiple models in the browser.

Three APIs make switching from existing tools easy

Sakura AI Engine accepts requests in the same format as the OpenAI and Anthropic APIs.

If your app already uses the OpenAI API, you can try it by swapping the base URL, API key, and model name. The official website also links to an article on using Sakura AI Engine models from Claude Code via the Messages API.

Available models

Standard models

Public preview models

Qwen3-0.6B-cpu, Phi-4-mini-instruct-cpu, Qwen3-VL-30B-A3B-Instruct, Kimi-K2.6, Qwen3.6-35B-A3B, gemma-4-31B-it, Kimi-K2.7-Code, Qwen3-Embedding-4B(FP16), and Weblab-MedLLM-gpt-oss-120b are offered on a trial basis.

Public preview is provided to collect real-world testing and feedback. Models may be discontinued without notice, quality is not guaranteed, and prices may change when they become generally available. Keep this in mind before building them into a production service.

Closed models (Japanese commercial models)

PFN's "PLaMo 3.0 Prime" and NEC's "cotomi v3" are also available, but require approval by the model provider.

• Apply from "Available models" in the control panel. Results usually take about 7 business days
• Only corporate accounts can apply (not available to personal accounts)
• Not covered by the free tier and not usable on the Foundation Model Free Plan. For pricing, check the manual site or contact the company
• Input data for inference is never sent to the model provider

Pricing

Standard edition: two plans and a free tier

The standard edition has two plans: the "Foundation Model Free Plan" and the "Pay-as-you-go Plan". Both include the same free tier, which resets automatically each month.

Plans do not switch automatically. Using up the free tier on the Free Plan never triggers charges, but you need to change to the Pay-as-you-go Plan to keep going. The company also states that the Free Plan has a cap on sign-ups and will stop accepting new users once it is reached.

Standard model pricing (tax included)

A mora is a unit of sound length equal to one kana character.

RAG alone has no free tier and is charged even on the Foundation Model Free Plan. The manual warns that charges continue from when you start using RAG until you delete all documents.

On the Pay-as-you-go Plan, usage beyond the free tier is charged per each model's minimum unit (10,000 tokens, or 60 seconds for transcription).

Public preview model pricing (tax included)

In the official price table, the free tier for the two standard text generation models, and the free tier for the seven public preview text generation models, are each shown in a single merged cell as "up to 3,000 requests per month". The table does not make clear whether this is per model or shared.

Private Edition pricing

Pricing is a flat monthly fee per GPU, quoted individually based on configuration. No prices are published. Quotes and consultations are free through the inquiry form.

What the Private Edition offers

The Private Edition is a fully managed service in which SAKURA internet hosts your chosen model on dedicated GPU infrastructure. SAKURA internet handles everything from building the inference platform to operations, monitoring, and incident response.

• Host the model you want Besides open-weight models such as Kimi and Gemma, you can entrust your own models or fine-tuned models. They are delivered ready to use via API.
• Flat pricing regardless of usage Charges do not vary with token usage, which makes budgeting easier for high-traffic services or services with hard-to-predict usage. As a domestic service, it is also unaffected by exchange rates.
• Dedicated GPUs and closed network connections Dedicated GPUs keep performance stable regardless of other users' load. Closed network connections that keep input data and model assets from leaving your environment are also supported.
• Input data is not used for training As with the standard edition, inputs and outputs are not stored or used for training (ZDR).
• Pin a model version Avoid changes in answer behavior caused by model updates by continuing to use a specific version.

The company gives three example setups: restricted use within the company or for specific customers, a platform for offering your own AI service externally, and combined use with external models from OpenAI or Anthropic.

Data handling

• Inputs and responses are not stored and are not used for model training or service improvement (ZDR)
• Prompts are not cached
• Inputs and responses are processed in data centers in Japan
• Even when using closed models, inference input data is not sent to the model provider

Pros and cons

Pros

• Runs entirely in Japanese data centers, with an explicit commitment not to store or train on input data
• Text generation is free for up to 3,000 requests per month. On the Free Plan you are never charged even if you exceed the limit
• OpenAI- and Anthropic-compatible APIs make it easy to switch from existing apps or tools such as Claude Code
• Text generation, transcription, embeddings, text-to-speech, and RAG all in one service
• Japanese commercial models such as PLaMo and cotomi are available through the same API upon application
• A path to the flat-rate, dedicated Private Edition as usage grows

Cons

• The standard edition is not covered by an SLA
• Even on the Free Plan, a 429 error (too many requests) may be returned temporarily during heavy traffic
• Model choice is limited to the provided list. Models from overseas commercial APIs such as GPT and Claude are not included
• Many models are in public preview and may be discontinued or repriced without notice
• RAG has no free tier and is charged even on the Free Plan
• Private Edition pricing is not published and requires an inquiry

Reputation and reviews

Judging from the official announcements and terms, Sakura AI Engine is evaluated along these lines.

• Likely reasons to choose it: It explicitly runs within Japan and does not store or train on data. It is easy to try on the free tier, and the Free Plan never charges you for going over the limit. The official website also links to guides for knowledge search with Dify and for integration with Claude Code.
• Points to watch when comparing: Only models on the provided list are available, and models from overseas commercial APIs such as GPT and Claude are not included. It suits uses that prioritize data handling and cost over cutting-edge performance.

Before adopting it, try your own prompts in the Playground and check response quality and speed.

FAQ

Q. Can I use Sakura AI Engine for free? A. Yes. On the Foundation Model Free Plan, text generation with models such as gpt-oss-120b is free for up to 3,000 requests per month. However, Sakura Cloud requires a registered credit card, and RAG is charged even on the Free Plan.

Q. What happens when I use up the free tier? A. On the Foundation Model Free Plan, rate limiting restricts your requests. It never switches to pay-as-you-go automatically. On the Pay-as-you-go Plan, usage beyond the free tier is charged.

Q. What is the difference between the standard edition and the Private Edition? A. The standard edition shares models prepared by SAKURA internet and charges pay-as-you-go. The Private Edition runs your chosen model on dedicated GPUs for a flat monthly fee per GPU. Closed network connections and bringing your own or fine-tuned models are supported in the Private Edition.

Q. How much does the Private Edition cost? A. Prices are not published. It is a flat monthly fee per GPU, quoted individually by configuration. Quotes are free.

Q. Is my input data used for training? A. No. The company states that inputs and responses are not stored or used for training or service improvement, and prompts are not cached.

Q. Can I use Japanese LLMs? A. Yes. llm-jp-3.1-8x13b-instruct4 is available as a standard model with a free tier. PFN's PLaMo 3.0 Prime and NEC's cotomi v3 are available to corporate accounts that apply and pass the provider's review.

Q. Can I use it with Claude Code? A. Yes, via the Anthropic-compatible Messages API. However, the only model currently available through the Messages API is "preview/Kimi-K2.6".

Q. Is anything limited to corporate accounts? A. Closed models (PLaMo 3.0 Prime, cotomi v3) can be applied for only by corporate accounts. They are not available to personal accounts.