Use · Sakura Internet AI Engine

Is it good enough?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D · 48/100

This page is a projection of the one entry record, the Reliability factor that Use covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Models served

Prices and the live list are in the console. The catalogue, with Sakura's stated licences, is as follows.

  • gpt-oss-120b and llm-jp-3.1-8x13b-instruct4 (Apache-2.0, not preview).
  • Qwen3.6-35B-A3B, Qwen3-VL-30B-A3B-Instruct, Qwen3-0.6B, Gemma 4 31B it and Weblab-MedLLM-gpt-oss-120b (Apache-2.0, preview).
  • Phi-4-mini-instruct (MIT, CPU, preview) and Kimi-K2.7-Code (Modified MIT, preview).
  • GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017 (MIT, preview, added 2026-11-18).
  • Whisper large-v3-turbo, multilingual-e5-large and Qwen3-Embedding-4B, and VOICEVOX voices for speech.
  • Closed models by application: PLaMo 3.0 Prime (Preferred Networks) and cotomi v3 (NEC), under proprietary terms.

Licences are Sakura's statements and are unchecked against each model's repository. A model asked for its own name may answer wrongly (manual).

Inference features

Streaming, JSON response format, tool calling (not on llm-jp, Phi-4-mini, Qwen3-0.6B, Qwen3-VL), vision on most preview models (three images per request on Qwen3-VL) and per-model reasoning control. gpt-oss-120b cannot disable reasoning, and on Kimi-K2.7-Code disabling it leaks reasoning into the answer.

Also available are embeddings, Whisper transcription (30 minutes or 30 MB per file), speech synthesis (about 1,000 characters), and a hosted RAG vector store with two embedding models and cosine or L2 search. Models whose names start with preview/ can change or be withdrawn without notice.

Use restrictions that come with the terms: no military use except for Japan's defence (AI Engine terms article 12(1)(vi)), no disinformation or discrimination uses, no use by export-listed parties.

Context and token limits

Context length per Sakura's table:

  • gpt-oss-120b: 131,072 tokens.
  • Gemma 4, Qwen3-VL, Qwen3.6 and Kimi-K2.7-Code: 262,144.
  • GLM-5.3-Flash and DeepSeek-V4.1-Flash-1017: 1,048,575.
  • Qwen3-0.6B: 40,960.
  • llm-jp-3.1 and Phi-4-mini: 4,096.

Maximum output is 32,000 tokens on most models. A request is cut off after about 5 minutes, so slow models (Gemma 4, Kimi-K2.7-Code) may stop before the token cap.

Fine-tuning and custom models

The shared service has no fine-tuning. Private Edition hosts a model the customer owns or has the right to use on a GPU platform dedicated to that customer, with Sakura running, monitoring and patching it.

Training, retraining, fine-tuning, RAG build, prompt tuning, guardrail design and closed-network connection are excluded unless written into the order form (article 2(4)). The customer is responsible for the model's safety, legality and guardrails (article 13). Private Edition is open to corporations established under Japanese law only (article 3). PLaMo forbids fine-tuning outright.

How this scores

The ownership factor this domain covers, drawn from the one entry record.

3

ReliabilityDoes it stay up and stay secure?

Moderate

Production service with a company-level security programme, but no SLA, most models in preview that can end without notice, and no AI Engine status history (manual 01-basics; terms Art. 11 of Private Edition).

How this scores (AOI sub-dimensions)
Reliability2/5whether it stays up, with an SLA and status historyServing is in production, but AI Engine is expressly outside the SLA (manual), Private Edition is best effort unless an SLA is written, the Sakura Cloud status page has no AI Engine component, and no incident history is published.
Security2/5the controls protecting your traffic and dataTemporarily stored inputs and outputs are encrypted (Art.
Compliance2/5which independent certifications and attestations it holdsSakura holds ISO 27001, ISO 27017, ISO 27018, ISMAP and Privacy Mark at company level, which would meet a score of 3 for the company.
What this means for adoptionInference here is partly owned. Sakura's terms commit in writing not to use inputs, not to give them to model providers and not to train on them, which is a real baseline. They do not promise zero retention, do not fix the data location, and come with no SLA or DPA. Ownership would move to substantial if Sakura wrote Japan-only processing and a bounded retention period into the terms, published a processing agreement, and confirmed AI Engine inside its ISMS scope.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Terms of serviceread2026-11-19
Sakura AI Engine service terms (さくらの AI Engine サービス約款, in force 2025-09-24).
Terms of serviceread2026-11-19
Sakura AI Engine Private Edition service terms (in force 2026-11-02).
Terms of serviceread2026-11-19
Sakura basic terms (基本約款, applied 2023-10-02).
Privacy Policyread2026-11-19
Sakura personal information protection policy (revised 2024-09-02): Privacy Mark and JIS Q 15001 system; no third-party provision of personal information without consent except to bound contractors, to protect life or property, or where law requires.
Securityread2026-11-19
Sakura security page lists the following.
Securityread2026-11-19
Sakura cloud service information security policy (2018-07-01, revised 2025-01-24): full virtualisation isolates each customer-created server's memory and network interface; Sakura does not monitor, edit or disclose customer data without permission except as its terms provide; two-step authentication for the control panel.
Securityread2026-11-19
Sakura information security check sheet PDF (updated 2026-11-13): certificate expiry dates (ISMS 2027-04-12, PCI DSS 2027-03-14, Privacy Mark 2028-07-10), policies, access control, periodic vulnerability scanning, annual internal audits, outsourcing of billing, invoicing and customer support to unnamed contractors with periodic security checks.
Securityread2026-11-19
Sakura Cloud ISMAP scope page (2026-06-30): registered service is Sakura Cloud IaaS in Tokyo zones 1 and 2 and Ishikari zones 1 and 2 (servers, storage, network, appliances).
Third-party analysisread2026-11-19
Independent auditor's SOC 3 report on Sakura's Ishikari data centre housing, colocation and common services: security and availability criteria, period 2025-06-01 to 2026-05-31, unqualified opinion dated 2026-07-31.
Third-party analysisunverified2026-11-19
Sakura states a SOC 2 Type 2 report (Ishikari data centre) and a SOC 2 Type 1 report (Sakura Cloud) exist and are disclosed to customers and prospects on request.
Data Processing Addendumunverified2026-11-19
No data processing agreement, GDPR addendum, sub-processor list or data-location statement appears on the terms index, the AI Engine product pages or the manual.
Securityread2026-11-19
Sakura transparency report page: reports on legal and other requests cover Rental Server, VPS, Sakura Cloud, Dedicated Server PHY, Domain and Housing.
Documentationread2026-11-19
The Sakura Cloud manual page on AI Engine service basics (updated 2026-11-11) says the service is outside quality assurance (SLA).
Documentationread2026-11-19
Sakura Cloud manual, chat model specification (2026-11-18): context length, max output, vision, tool-call and reasoning support per model (4,096 up to 1,048,575 tokens); OpenAI-compatible parameters, unsupported parameters ignored; about 5 minutes per request; preview models can change or end without notice.
Documentationread2026-11-19
The Sakura Cloud manual operation guide (2026-11-11) and Tips (2026-05-14) cover the RAG API and vector store, transcription limits (30 minutes or 30 MB) and speech limits.
Documentationread2026-11-19
The Sakura Cloud manual closed-models page (2025-12-02) says PLaMo and cotomi need an application that the model provider approves (about 7 business days), are outside the free plan and are for corporate accounts only.
Documentationread2026-11-19
Sakura Cloud manual, licences of served models (2026-11-18): gpt-oss-120b, llm-jp-3.1, Qwen3 family, Gemma 4 and MedLLM under Apache-2.0; Phi-4, e5, Whisper, GLM-5.3-Flash and DeepSeek-V4.1-Flash under MIT; Kimi-K2.7-Code under Modified MIT; PLaMo and cotomi under proprietary end-user terms.
Licenceread2026-11-19
PLaMo licence agreement (closed model via Sakura): use only inside AI Engine, no fine-tuning or quantisation (Art.
Licenceread2026-11-19
cotomi v3 usage conditions (closed model via Sakura): internal business use only, no hosting or SaaS to third parties without consent (Art.
Vendor announcementread2026-11-19
The Sakura AI Engine product page says 'input data is not trained on or stored', 'ZDR (Zero Data Retention) compatible' and 'model execution and communication all completed in domestic cloud'.
Vendor announcementread2026-11-19
Sakura AI top page FAQ: 'model execution, communication and storage are all completed on cloud infrastructure in Japan' and 'can be operated under Japanese law'.
Vendor announcementread2026-11-19
Sakura AI Engine Private Edition page: dedicated environment, closed-network connection supported, ZDR, flat monthly fee per GPU by quote, bring-your-own or fine-tuned models.
Slaread2026-11-19
The Sakura Cloud service status page (status.cloud.sakura.ad.jp) lists Cloud API per zone (Tokyo 1 and 2, Ishikari 1 to 3), websites, DNS, GSLB and load balancer appliances.
Documentationread2026-11-19
Sakura Cloud incident and maintenance notice policy (2025-02-20): announcement channels and criteria for Sakura Cloud resource types.
Documentationread2026-11-19
Sakura Cloud support page (2026-09-29): standard support free on weekdays 10:00 to 18:00 JST by e-mail form or reserved call-back; premium paid support 24 hours, first reply within 2 hours for faults.