A NAS that answers questions: the QNAP QAI-h1290FX explained

Most offices keep their contracts, SOPs and reports on a file server, and nobody can find anything in it. The QNAP QAI-h1290FX adds a graphics card to that server so people can ask a question in plain language and get the answer, with the file it came from.

It is a NAS (network-attached storage, a shared file server) with twelve SSD bays and one NVIDIA professional card inside. The files and the AI live in the same box, so nothing has to be uploaded anywhere.

What is inside the box

QNAP lists these specifications for the model we offer:

• Processor and memory: AMD EPYC 7302P (16 cores, 32 threads) and 128 GB of ECC memory, expandable to 1 TB.
• Storage: twelve 2.5-inch bays for U.2 NVMe or SATA SSDs. No hard drives. It runs QuTS hero, QNAP's ZFS-based system, so snapshots let you roll back a deleted or overwritten file.
• Network: two 25GbE ports and two 2.5GbE ports.
• Graphics: one NVIDIA RTX PRO Blackwell card, 32 GB or 96 GB.

The AI software runs as containers (apps in their own sandbox) on the same machine. QNAP provides templates for Ollama, AnythingLLM, Open WebUI, ComfyUI, Stable Diffusion, n8n and Whisper, and says the box also supports the vLLM inference engine for serving many users.

How the box answers from your files: RAG

A language model only knows what it learned in training. It has never seen your contracts. RAG (retrieval-augmented generation) fixes that without retraining anything. In three steps:

1. Index. The box reads your documents once and turns passages into embeddings: lists of numbers that capture meaning, so a question about "notice period" lands near the clause that says "60 days written notice". These go into a search index. QNAP's file search, Qsirch, provides this part.
2. Retrieve. When someone asks a question, the index returns the few passages that match best.
3. Generate. Those passages are handed to the LLM (large language model) along with the question. The model writes the answer from them and names the files it used.

That last point is why this is not just a chatbot. The model is working from text in front of it, so you can open the cited file and check. It can still make mistakes, so for contracts the file is the authority.

Access matters as much as answers. Decide which folders each team's assistant may index before you load anything confidential, so finance files never feed a sales chat. Users and folder access are part of the setup.

What staff actually do with it

• Ask the files. "What are the return conditions in the warehouse SOP?" The answer comes in a few lines with the SOP attached.
• Summarise and draft. Turn a long report into five points for the board, in formal Bahasa Indonesia or English. A model family such as SEA-LION, trained for Southeast Asian languages, helps here.
• Meeting minutes. Whisper (speech-to-text) transcribes a recording, then a model writes minutes and a task list per person.
• Routine automation. n8n, a workflow tool, can read new purchase-order emails every morning and fill a table. An AI step in the workflow does the reading.

Why only two cards fit

You might expect to drop in any GeForce card. You cannot. QNAP's specifications limit the card size in slots 3 and 4 to 280 mm long and 111.15 mm tall, and the power supply is 750 W. The card shares that supply with the processor, memory and twelve SSDs.

QNAP validates two cards: the NVIDIA RTX PRO 4500 Blackwell (32 GB, 200 W) and the RTX PRO 6000 Blackwell Max-Q (96 GB, 300 W). The full-power RTX PRO 6000 Workstation Edition is a 600 W card, so the 300 W Max-Q version is the one for this box. GeForce cards such as the RTX 5080 and 5090 are longer and draw more, and QNAP does not list them. QNAP lists no two-card setup, so plan on one card per box.

What each package runs

We sell the box in two packages: Office AI with the 32 GB card and Company AI with the 96 GB card. The table shows five models. Q4 means a 4-bit file, a compressed model about a third of the original size; Q8 is 8-bit, closer to the original but larger. Speed is tok/s (tokens per second, roughly words per second) for one user at a 32,000-token context, the size of a long document. "Chats" is how many 32K conversations fit in memory at once.

The fast figures come from MoE (mixture-of-experts) models, which only use a small part of their weights for each word. Dense models such as Qwen3.8 27B use all of them and run slower. The conversation counts are a memory limit; replies slow down when many people type at the same moment.

Who it suits

This box suits an office that needs a file server anyway and wants AI on top of it: law and notary firms, distributors, clinics, production houses. Storage comes first, and the AI is a strong second job. If the AI is the main job and you need much more graphics memory, a GPU tower is the better tool. Our AI workstation preset, with two RTX PRO 6000 Max-Q cards, has 192 GB; it is the alternative for that case.

Open models are strong at document questions and drafting, but the best cloud models still reason better on hard problems. Many companies use a cloud service for general writing and keep confidential files on the box. Cloud AI is often cheaper per answer for light use; the box is a fixed cost with no per-seat fee.

The short version

• The QAI-h1290FX is a 12-bay all-flash NAS with one RTX PRO card, so the AI reads files where they already live.
• RAG means: index your files, retrieve the matching passages, let the model write the answer and cite the file.
• Only the RTX PRO 4500 (32 GB) and RTX PRO 6000 Max-Q (96 GB) fit, because of the 280 mm slot limit and the 750 W supply.
• Pick it when storage and private document Q&A matter most. Pick a GPU tower when you need more graphics memory.

All speeds and conversation counts are estimates. See the QNAP AI server page for the full model list, or try other machines in the hardware finder.