What Is Open GENAI? Japan's Digital Agency Open-Sources Its Government AI Platform hirokawaguchi/open-genai Brings Full Local Deployment (Keycloak / Ollama / Qdrant / Stable Diffusion / faster-whisper)

Open GENAI is the open-source release of GENAI (源内), Japan's government generative-AI platform — published by the Digital Agency of Japan on April 24, 2026 on GitHub under MIT license.

Three takeaways:

1. A nation's AI platform is now free and commercially reusable — MIT + CC BY 4.0, fully modifiable and redistributable 2. Multi-cloud (AWS / Azure / Google Cloud) — REST API-first + ExApp microservices unify the cloud-of-choice behind one interface, structurally eliminating vendor lock-in 3. A 180,000-person pilot in 2026 — covering every government ministry, with expansion targeted to municipalities and the private sector

This column sits next to our PLaMo 3.0 Prime (a model selected for GENAI's trial) and Local LLM June 2026 Update as part of the Japan public-sector AI cluster.

"Open GENAI" is the popular nickname for the OSS-released subset; the official designation is the OSS portion of GENAI.

1. [GENAI Web](https://github.com/digital-go-jp/genai-web) — full-stack AI interface:

- TypeScript 6 + React 19 - Zustand 5 (state) - React Router 7 - AWS CDK (infra-as-code) - Tailwind CSS + Digital Agency Design System — government-standard UI - AWS deploy walkthrough included; ~20 minutes for 254 resources

2. [GENAI AI Apps](https://github.com/digital-go-jp/genai-ai-api) — three cloud-specific development templates:

- AWS: administrative RAG template (Bedrock + Knowledge Base lineage) - Microsoft Azure: self-hosted LLM template (open-weights models on Azure) - Google Cloud: legal-system AI — answers grounded in current statutes

Validated models: Claude Sonnet 4.6, Amazon Nova Lite. [PLaMo 3.0 Prime](../columns/plamo-3-0-prime-pfn-japanese-llm-2026-06) has been selected as a trial model in GENAI.

The headline design choice is "REST-API-first + ExApp (external app integration)".

REST API-first: GENAI Web calls every backend feature via REST. Whether the backend lives on AWS, Azure, or Google Cloud doesn't matter — same API spec, same UI.

ExApp (External App Integration): features that don't belong in the core (video generation, web extraction, prompt optimization) are bolted on as independent microservices. Core stays minimal; extensions live outside.

Vendor lock-in avoidance: nothing in the design forces a single cloud, so each agency or municipality can use its own procurement-permitted cloud. That's the essential requirement for a government AI platform.

GENAI shares an AWS-CDK foundation with AWS's official GenU, but their design philosophies diverge significantly.

The summary: GENAI is "a multi-cloud, minimal, governance-heavy base for public-sector and enterprise use"; GenU is "a developer-oriented, AWS-managed, feature-rich reference implementation."

Prerequisites:

- CDK bootstrap in two regions — `ap-northeast-1` (Tokyo) and `us-east-1` (CloudFront WAF requirement) - Sufficient AWS account permissions (IAM / CloudFormation / Bedrock, etc.)

Flow:

1. CDK deploy (~20 min, 254 resources) 2. User sign-up via CloudFront URL 3. Run `add-system-admin.sh` to enable team management 4. Run `create-common-app-team.sh` to expose generic AI apps

Unlike GenU (which works fully right after deploy), GENAI requires administrative scripts — a governance-first design choice.

The Digital Agency's official release targets AWS / Azure / Google Cloud. [hirokawaguchi/open-genai](https://github.com/hirokawaguchi/open-genai) is an unofficial community fork by Hiro Kawaguchi that runs GENAI entirely on local infrastructure. MIT-licensed, experimental, 68 stars / 4 forks (as of writing), TypeScript 93.3% + Python 5.7%.

The README explicitly states "non-official fork, no affiliation with the Digital Agency." It absorbs the Digital Agency's `genai-web` with modifications; the `genai-ai-api` component is obtained separately. Original licensing and attribution are preserved in `genai-web/LICENSE`.

The defining choice is that every cloud dependency is replaced with an OSS local alternative:

Everything runs on OSS — zero dependency on AWS / Azure / GCP, viable for air-gapped environments and on-prem-only deployments.

One Docker Compose command spins up the full set of services:

- Web frontend (React / Vite) - FastAPI backend - RAG microservice - Whisper transcription app (faster-whisper) - Stable Diffusion proxy (`sd-app`) - Dify integration layer — workflows / chatflows pluggable as ExApps - Keycloak identity provider - Qdrant vector DB

Each microservice has health checks, and optional services can be opted in. Local extraction of PDF / DOCX / XLSX / TXT is included.

LLM compatibility: any server exposing the OpenAI-compatible `/v1` endpoint works.

- Ollama (default) — ideal for individual PoC / SMB - vLLM — GPU-server production - LM Studio — desktop GUI - OpenAI API — pure cloud fallback (within license terms)

Project recommendation: Qwen2.5 for Japanese-language quality. Swap freely with Llama 3.x, Gemma 4, Liquid AI LFM2.5-J, PLaMo 3.0 Prime, etc.

Hardware / platform:

- macOS Apple Silicon: Ollama and Stable Diffusion run on the host (Metal); containers don't see the GPU directly - Linux + NVIDIA: containers can use CUDA via the NVIDIA Container Toolkit — production-grade - CPU-only: works, but LLM inference is heavy (PoC / small-scale only)

A single MacBook Pro M3 can realistically run the full GENAI stack for evaluation.

Dify integration: Dify (the OSS LLM-app platform) workflows / chatflows can be plugged in as ExApps. Design complex business processes visually in Dify and call them from the GENAI UI.

Team-scoped knowledge bases: Qdrant collections are isolated per team, so multiple departments' RAG data don't mix.

SAML auth: Keycloak-based — integrate with your corporate IdP (Microsoft Entra ID / Okta / Google Workspace etc.) for standard auth flows even in closed organizations.

The two are complementary: production runs on the official release; evaluation, small-scale, and confidentiality-sensitive work runs on the local fork. For data-sovereignty-first engagements (medical, defense, regulated finance), the local fork is sometimes the only viable option.

- Unofficial, experimental — spec changes and project abandonment are real risks; treat carefully in commercial production - No legal support — community-run by Hiro Kawaguchi; no SLA or formal accountability - `genai-ai-api` not bundled — fetch separately from `digital-go-jp/genai-ai-api` if needed - GPU memory requirements — comfortable Qwen2.5 inference wants 16–32GB VRAM (or equivalent Apple-Silicon unified memory) - Upstream maintenance burden — keeping in sync with Digital Agency's `genai-web` upstream falls on the operator

In fiscal year 2026, the Digital Agency rolls GENAI to ~180,000 government employees across every ministry. This means:

- Standardized generative-AI usage across the Japanese government - Production feedback at 180,000-person scale - Combined with OSS release, transferable lessons for municipalities and private adopters

The result is "the AI platform the government picked" branding + 180K-person operational evidence — a powerful procurement signal for municipalities and private buyers in 2027 onward.

PLaMo 3.0 Prime's selection as a GENAI trial model fits this same trend — Japanese from-scratch LLM + Japanese government OSS platform is shaping up to be the new default for Japan's public-sector AI.

1. Municipalities (prefectures and cities):

- Municipalities looking to avoid single-vendor lock-in - Those wanting to customize for local administrative needs - Those needing to meet government-standard UI / security requirements

2. Private companies (especially public-sector SI vendors):

- Use it as a baseline for their own products (fork + customize) - Pitch to government tenders with "based on GENAI" procurement language - Sell custom features as ExApps

3. SMBs and startups:

- Skip building governance features (RBAC / SAML / WAF / TTL) from scratch - MIT allows full commercial use - A solid base for SaaS targeting public-sector customers

4. Enterprise internal AI platform teams:

- A reference implementation for building multi-cloud AI platforms in-house - The Digital Agency design system as a starting point for internal UI standards

1. No permanent maintenance: the Digital Agency explicitly notes "permanent maintenance is not guaranteed and OSS publication may be terminated." Long-term operation is the adopter's responsibility. This is honest scope-setting from the government, but production adopters need either in-house or third-party maintenance.

2. Only "a part" is published: the released code is a subset of GENAI, not the full internal platform — internal versions may differ.

3. ExApp ecosystem still young: many features depend on ExApps, so practical use requires building or sourcing those external services. Until the ecosystem matures, expect to use GENAI as a minimal base.

4. AWS-centric base: the multi-cloud claim mostly applies to AI app templates; the Web stack and CDK infrastructure assume AWS.

5. Governance features have operating cost: SAML / KMS CMEK / multi-layer WAF / Bedrock Inference Profiles require ops know-how. For SMBs / startups that just want "AI on a webpage," the governance overhead may be more than they need.

We can introduce Open GENAI to customers through our AI consulting / software development practices in three patterns:

Pattern 1: Baseline for municipality / public-sector engagements

For projects in Shinagawa, Minato, Shibuya, Setagaya, Meguro, Ota — Oflight's home wards — fork GENAI, customize for local administrative work (citizen records, tax, education FAQ), and pitch with the powerful "based on the Digital Agency's GENAI" procurement story.

Pattern 2: Reference implementation for mid-market enterprises

For mid-market enterprise internal AI platforms, fork GENAI rather than starting from zero, simplify to fit the business, and add functionality via ExApps. Default-built-in governance (RBAC / SAML / KMS / WAF) is the biggest value-add for enterprise buyers.

Pattern 3: ExApp custom development

As the GENAI ecosystem grows, an ExApp custom-development market is taking shape. Video generation, web extraction, OCR, translation, domain-specific search — Oflight can build and deliver ExApps for organizations needing specific functionality.

We help municipalities and private companies adopt and customize Open GENAI.

"Does GENAI fit our agency / company?" "Which ExApps should we build?" "What's the multi-cloud strategy?" — 1–2 weeks, written report.

👉 Contact us — AI consulting

Fork GENAI and customize for municipal or enterprise operations + ExApp development. Engineers experienced with CDK / TypeScript / React 19 deliver the work. Full SI engagements typically start around ¥5,000,000+ (industry benchmark).

👉 Contact us — custom implementation / SI

We cover the gap left by the Digital Agency's no-permanent-maintenance stance — upstream-update tracking, security patching, LLM model swaps, ExApp additions on a monthly retainer.

- [OpenClaw maintenance](../services/openclaw-setup): Light ¥9,800 / Standard ¥19,800 / Premium ¥49,800 per month - AI-consulting continuous support: Light ¥30,000 / Standard ¥80,000 / Premium on request

👉 Contact us — ongoing maintenance

Q1. Is commercial use allowed? A. Yes. MIT + CC BY 4.0 — commercial use, modification, and redistribution are all free. Embed it in products, ship it as SaaS, do whatever you like. Q2. GENAI or GenU — which one? A. GenU if you want AWS managed features fast, with a developer-prototype mindset, AWS-only. GENAI if you need public-sector / enterprise governance, multi-cloud support, and long-term operation. Q3. How serious is the "no permanent maintenance" caveat? A. Production adoption requires either in-house or third-party maintenance. The Digital Agency will continue updates for now but won't promise forever. This is standard OSS reality — Linux and Kubernetes upstreams don't guarantee forever either. Our monthly maintenance plans cover this gap. Q4. Why combine GENAI with PLaMo 3.0 Prime? A. Data sovereignty + Japanese from-scratch model + government-standard UI all in one stack. See our PLaMo 3.0 Prime column. A practical pattern: overseas models (Claude / GPT) for general work, sensitive workloads on Japanese model + Japanese OSS platform. Q5. What's an ExApp? A. A way to bolt on features not in core GENAI — video generation, web extraction, OCR, translation — as independent microservices. They speak to GENAI via REST API. Each organization adds only the ExApps it needs. Q6. Will the 180,000-person pilot results be published? A. Expected in the second half of FY2026. Feedback should flow back into the OSS. Interim updates via the Digital Agency's official site. Q7. Why hire Oflight? A. We're strong on AWS CDK / TypeScript / React / public-sector workflows, and we deliver GENAI fork → customize → ExApp → monthly maintenance end-to-end. Particularly strong in Shinagawa / Minato / Shibuya / Setagaya / Meguro / Ota — our home wards. Q8. Comparable initiatives in other countries? A. Singapore's Pair / Launchpad, the US's AI.gov, the UK's i.AI are analogous. GENAI's distinguishing feature internationally is multi-cloud + OSS — a model worth watching as a Japanese contribution to public-sector AI.

Open GENAI (the OSS release of GENAI) is Japan's national government generative-AI platform, published on April 24, 2026 by the Digital Agency under MIT license.

Three essential value props:

1. A national AI platform, free and commercial-OK — structurally lowering generative-AI platform costs for municipalities and private companies 2. Multi-cloud + built-in governance — REST API + ExApp + RBAC / SAML / KMS / WAF / TTL satisfies enterprise requirements by default 3. 180K-person pilot + Japanese-model trials (PLaMo etc.) — "the AI platform the government picked" carries serious procurement weight

Reservations: no permanent maintenance, only "part of" GENAI is published, ExApp ecosystem still young, AWS-centric base, governance operations have real cost.

Oflight recommendations:

- Municipal / public-sector engagements: a GENAI-based proposal is hard to beat - Mid-market enterprises: a strong reference implementation with governance baked in - SMBs / startups: zero-cost baseline + commercial freedom for fast AI platform build-out - Ongoing maintenance: fill the Digital Agency's no-forever-support gap with our monthly retainers

Primary: - Digital Agency announcement (English) - GitHub: digital-go-jp/genai-web - GitHub: digital-go-jp/genai-ai-api - Digital Agency official Japanese coverage: - INTERNET Watch — Digital Agency open-sources GENAI - Ledge.ai — GENAI OSS released, commercial use allowed - ITmedia AI+ — Government AI GENAI open-sourced - SBBit — Digital Agency open-sources GENAI - CodeZine — GENAI GitHub publication starts - TECH NOISY — Supporting municipal AI infrastructure - Innovatopia — Breaking vendor lock-in by design - Mμgen — OSS release accelerates AI adoption - Parse — SMB digital transformation perspective Technical: - Classmethod DevelopersIO — GENAI vs GenU diff + AWS deploy - Qiita (trailfusion_ai) — 5-minute three-cloud guide - note (makokon) — individual perspective Related Oflight columns: - PLaMo 3.0 Prime — trial model in GENAI - Local LLM June 2026 Update - Sakana Fugu — orchestration model - Liquid AI LFM2.5-J — Japanese-specialized models - Claude Fable 5 export-control suspension - Loop Engineering - Ornith-1.0 - Grok Build Oflight services: - AI Consulting - Software Development - OpenClaw Setup Inquiries: - Evaluation & PoC (AI consulting) - Custom implementation / SI (software development) - Ongoing maintenance (OpenClaw maintenance)

Local fork (community): - GitHub: hirokawaguchi/open-genai (unofficial local-deployment fork) - GitHub: hirokawaguchi