Ask your files
They typeApa syarat retur barang dari reseller menurut SOP gudang?
They getThe answer in a few lines, with the SOP file it came from.
Runs on QNAP Qsirch RAG search or AnythingLLM, with a local model.
Home / Servers / QNAP AI server
QNAP QAI-h1290FX private AI server
Ask your company files anything.
Nothing leaves the office.
One box in your server room keeps your files and runs the AI. Your team asks in Bahasa Indonesia or English, gets the answer with the file it came from, and only searches folders they are already allowed to open.
- Runs on its own NVIDIA card. No cloud account and no monthly fee per user.
- Also your file server: twelve SSD bays with ZFS snapshots.
- Set up by EMD Lab, with online support after delivery.
Mencari di: Legal › Kontrak
Berapa lama masa pemberitahuan untuk mengakhiri kontrak dengan PT Sinar Jaya Abadi?
60 hari kalender sebelum tanggal berakhir, disampaikan secara tertulis (Pasal 14.2). Addendum 2025 tidak mengubah pasal ini.
Kontrak Distribusi Sinar Jaya 2024.pdf
Addendum 01 - 2025.pdf
Hanya dari dokumen yang boleh Anda buka
Tanya apa saja tentang dokumen kantor…
Your team already uses AI. The question is where your files go.
Contracts pasted into public chatbots
Staff copy customer data, prices and contract clauses into free AI tools to work faster. From then on, that text is kept under the provider's rules, not yours.
Answers buried in folders
Finding one clause or one SOP step means opening file after file, or waiting for the one person who remembers where it is.
Files spread everywhere
Laptops, email attachments, flash drives and an old NAS. There is no single place to search, back up or protect.
One box solves all three
- Your files in one place. Up to twelve SSDs, with ZFS snapshots to roll back mistakes or ransomware.
- A private AI. An NVIDIA RTX PRO card inside the box runs the models. No cloud account.
- Search that follows your folder permissions. People only get answers from files they can already open (QNAP's Qsirch RAG search).
What your team types, and what they get back
Four jobs it does from the first week. The examples are in Bahasa Indonesia because that is how most teams will use it; English works the same way.
Summarise and draft
They typeRingkas laporan audit 40 halaman ini jadi 5 poin untuk direksi, bahasa formal.
They getFive points in formal Indonesian, ready to paste into an email and edit.
Runs on Open WebUI with SEA-LION, a model family trained for Indonesian and other Southeast Asian languages.
Turn meetings into minutes
They typeUpload rapat-mingguan.mp3, lalu: buat notulen dan daftar tugas per orang.
They getA transcript, the minutes and a task list per person. A one-hour recording takes minutes, not hours.
Runs on Whisper speech-to-text, which handles Indonesian.
Automate routine work
Set up onceSetiap pagi: baca email PO baru, ambil nomor PO dan tanggal kirim, masukkan ke tabel.
They getA workflow that runs by itself every morning, inside your office network.
Runs on n8n workflows. EMD Lab sets up the first ones with you.
It also runs image generation for marketing (ComfyUI, Stable Diffusion) and a private coding assistant for your IT team.
Two packages. Pick by team size.
QNAP lists two NVIDIA cards for this machine. The card decides how many people can use it at once and how large the AI models can be.
RTX PRO 4500 Blackwell
32 GB graphics memory, 200 W
For teams of up to about 30 people.
- Main assistant
- Nemotron SEA-LION 30B, for Indonesian and other SEA languages
- Conversations at once
- About 15 long or 39 short
- Answers appear
- Faster than anyone can read them
- Largest models
- About 30 billion parameters
- Also runs
- Meeting transcription, image generation, a coding assistant
RTX PRO 6000 Blackwell Max-Q
96 GB graphics memory, 300 W
For about 30 to 100 people, or when you need 120B-class models.
- Main assistant
- gpt-oss-120b, OpenAI's open-weight model
- Conversations at once
- About 23 long or 94 short
- Answers appear
- Faster than anyone can read them
- Largest models
- About 120 billion parameters
- Also runs
- Everything in Office AI, with room for larger models
Both packages include
- SSDs sized to your data
- Models and apps installed
- Users and folder permissions set up
- Stress-tested, with a written test report
- Online support after delivery
- 5-year QNAP warranty on the box (claims handled by us)
EMD Lab, backed by 25 years of distributing IT hardware in Indonesia, plans, builds and tests it before delivery, then supports it online, with an optional monthly care plan. Warranty claims go through us.
Estimates from the EMD Lab AI selector, not measurements. “Long” is a conversation of about 32,000 tokens, roughly a 40-page document; “short” is about 8,000. The counts are limited by graphics memory; replies slow down as more people type at the same moment. Team sizes assume only some of your staff use it at the same time. The graphics card carries its maker's warranty. Price on request.
See all 11 models and their estimated speeds
| Model | Indonesian | With RTX PRO 4500 (32 GB) | With RTX PRO 6000 (96 GB) | Licence |
|---|---|---|---|---|
| Nemotron SEA-LION v4.8 30B-A3BIndonesian and SEA-language chat, fast | Built for Indonesian | 4-bitabout 145 tok/s | 8-bitabout 176 tok/s | Check licence terms |
| Qwen SEA-LION v4.5 27BIndonesian assistant, document Q&A | Built for Indonesian | 4-bitabout 36 tok/s | 8-bitabout 42 tok/s | Free for business use |
| Qwen3.8 27BOffice assistant, RAG, coding help | Official | 4-bitabout 35 tok/s | 8-bitabout 42 tok/s | Free for business use |
| Gemma 4 26B-A4BFast assistant, reads images | Good | 4-bitabout 132 tok/s | 8-bitabout 159 tok/s | Free for business use |
| Gemma 4 31BDocument Q&A, reads images | Good | 4-bitabout 30 tok/s | 8-bitabout 37 tok/s | Free for business use |
| Qwen3-Coder 30B-A3BPrivate coding assistant | Basic | 4-bitabout 100 tok/s | 8-bitabout 133 tok/s | Free for business use |
| gpt-oss-20bFast general assistant | Basic | 8-bitabout 145 tok/s | 8-bitabout 200 tok/s | Free for business use |
| gpt-oss-120bStronger reasoning for many users | Basic | Needs the 96 GB card | 8-bitabout 175 tok/s | Free for business use |
| Qwen3.5 122B-A10BLarge-model quality, multilingual | Good | Needs the 96 GB card | 4-bitabout 129 tok/s | Free for business use |
| Nemotron 3 Super 120B-A12BAgents and tool use | Basic | Needs the 96 GB card | 4-bitabout 112 tok/s | Free for business use |
| Llama 3.3 70BClassic 70B dense model | Basic | Needs the 96 GB card | 4-bitabout 26 tok/s | Check licence terms |
Speed for one person at a 32,000-token context, with the whole model in graphics memory. “Free for business use” means the licence allows commercial use (Apache 2.0, MIT or similar); “Check licence terms” means it has conditions, and we go through them with you. To compare other machines, see Which hardware runs which AI?
Configure yours
The processor, memory and chassis are fixed. Choose the graphics card and the drives, and we confirm everything in a written quote.
Fixed in every box
- Processor
- AMD EPYC 7302P, 16 cores / 32 threads, up to 3.3 GHz
- Memory
- 128 GB DDR4 ECC RDIMM
- Chassis
- QNAP QAI-h1290FX tower, 750 W power supply
- Network
- 2 × 25GbE SFP28, 2 × 2.5GbE
- Operating system
- QuTS hero (ZFS)
- Warranty
- 5 years (QNAP)
Cloud AI or your own box?
Both have a place. This is how they differ for an office.
| Topic | Cloud AI subscription | QAI-h1290FX in your office |
|---|---|---|
| Your documents | Uploaded to the provider's servers | Stay on the box in your office |
| Who sees what | Depends on what each person uploads | Follows your folder permissions |
| Quality of the AI | The strongest models available | Strong open models, behind the best cloud models on hard reasoning |
| Cost | A fee per user, every month | One purchase, plus an optional care plan |
| Without internet | Stops working | Keeps answering |
| Your file server | Still needed separately | Included: 12 SSD bays with snapshots |
| Getting started | Today | After delivery and set-up |
Many companies use both: a cloud service for general writing, and the box for anything confidential.
Questions buyers ask
Is it as smart as ChatGPT?
No. The open models it runs are strong at document questions, summaries and drafting in Bahasa Indonesia, but the best cloud models still reason better on hard problems. If you need those, keep a cloud subscription for general work and keep confidential files on the box.
Does it need the internet?
Not for answers. The models run on the card inside the box. It uses the internet for updates and for our online support, and it is never left open to the internet.
Can we start small and upgrade later?
Yes. Start with Office AI; the card can be changed to the RTX PRO 6000 later, and memory grows from 128 GB to 1 TB. Plan on one graphics card: QNAP does not list a two-card setup for this model.
We already have a QNAP NAS. Is it wasted?
No. Keep it as the backup target for the new box. Older NAS models lack the graphics card these AI models need, so they cannot run them.
Do we need a special network switch?
The two 2.5GbE ports work with a normal office switch. The 25GbE ports need an SFP28 switch or direct cables; most offices only need them for video or large design files.
Specifications
- Model
- QAI-h1290FX-7302P-128G
- Processor
- AMD EPYC 7302P, 16 cores / 32 threads, up to 3.3 GHz
- Memory
- 128 GB DDR4 ECC RDIMM installed, 8 slots, up to 1 TB
- Drive bays
- 12 × 2.5-inch U.2 NVMe PCIe Gen4 or SATA SSD, shipped without drives
- Expansion
- 4 × PCIe Gen 4 slots (x16, x16, x8, x16)
- Graphics
- One NVIDIA RTX PRO Blackwell card: RTX PRO 4500 (32 GB, 200 W) or RTX PRO 6000 Max-Q (96 GB, 300 W)
- Network
- 2 × 25GbE SFP28, 2 × 2.5GbE
- USB
- 3 × USB 3.2 Gen 1
- Power supply
- 750 W, 100–240 V
- Size and weight
- 150 × 368 × 362 mm (H × W × D), 10.4 kg
- Operating temperature
- 0–40 °C
- Operating system
- QuTS hero (ZFS)
- AI apps
- QNAP templates for Ollama, AnythingLLM, Open WebUI, ComfyUI, Stable Diffusion, n8n and Whisper; vLLM for many users
- Warranty
- 5 years
Specifications from QNAP's product page and specification page, 23 October 2026. Availability in Indonesia and price on request.
Tell us about your team and your files
Send us how many people would use it, what files you keep and roughly how much data. We size the box, set it up with your folders and permissions, and send a written quote.
We reply with a written, itemised proposal.