512GB max RAM · 15 bays · 2 x 10GbE
AINAS Score 74
No verified price

AI ReadyQNAP
Why AI Ready: AMD EPYC 7302P or 7232P/7252 (no NPU, no iGPU) with PCIe Gen4 expansion, a GPU path of a free x8 slot where QNAP lists only the 6GB RTX 3050 LP, and a 4TB DDR4 ECC ceiling; QNAP documents no GPU inference on this model, so ready rather than native.
Updated
Can the QNAP TS-h2490FU run local AI?
Only lightly. Expect 70B class models at higher-quality quantization, or large mixture-of-experts models at 4-bit (Estimated) on the QNAP TS-h2490FU: AMD EPYC 7302P (16-core/32-thread, up to 3.3 GHz); AMD EPYC 7232P/7252 (8C/16T, up to 3.2 GHz) on 7232P models does the work with up to 4096GB of memory, which is fine for experiments rather than a daily assistant.
Its PCIe x8 slot leaves room to add a GPU when you outgrow CPU inference.
Local LLM support is documented by the manufacturer.
The TS-h2490FU suits buyers who want an all-flash rackmount NAS that can also host local AI software on its CPU. It has 24 bays, an AMD EPYC 7302P with 16 cores, 128GB of DDR4 ECC RDIMM memory as standard, and a 4096GB memory ceiling. Memory is upgradeable. Container Station supports Docker and the system supports virtual machines. Ollama and Open WebUI are vendor confirmed. Models up to Qwen3 235B A22B fit at Q4 (Estimated). It has 4 x 25GbE and 2 x 2.5GbE for fast network access, and a 5-year warranty. The AINAS Score is 72.
The main limit is the GPU path. QNAP lists one card for this system, the 6GB RTX 3050 LP, and it is low-profile only. No listed card is flagged as AI recommended. Local LLM support is marked limited, with mostly CPU inference. Buyers who want GPU-accelerated inference, or need 16GB of VRAM for Qwen's AI Mode in Qsirch, should look for a system with a full-height PCIe x16 slot, a larger GPU on the vendor's compatibility list, or an NPU. This system has no NPU and no integrated GPU, so a different build suits GPU-led AI work better.
No verified price
We have not verified a listing for the QNAP TS-h2490FU at a retailer we track. The specifications below stand on their own; a price appears here as soon as a listing is confirmed.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon.co.uk or Amazon.com, as applicable, at the time of purchase will apply to the purchase of this product.
Each capability shows how it was confirmed: Verified by AINAS, Vendor confirmed, Community confirmed or Unverified.
| Capability | Supported | Evidence | Note |
|---|---|---|---|
| Local LLM | Limited | Vendor confirmed | QNAP Ollama guide covers x86 QTS/QuTS hero but names no model. Strong CPU and RAM; only GPU on the list is the 6GB RTX 3050 LP, so mostly CPU inference. |
| Ollama | Yes | Vendor confirmed | Container Station Ollama tutorial for x86 QTS/QuTS hero NAS, with NVIDIA option. Model not named. |
| Open WebUI | Yes | Vendor confirmed | QNAP tutorial needs QTS 5.2 / QuTS hero h5.2 or later and Container Station 3. Model not named. |
| RAG | Limited | Vendor confirmed | Qsirch RAG Search needs 64-bit x86, QTS/QuTS hero 5.0.1+. Cloud RAG needs an OpenAI-compatible API; on-prem needs a compatibility-tested GPU. Model not named. QNAP lists this model as raw data storage for AI: https://www.qnap.com/en/solution/optimize-ai-with-qnap-storage |
| AI search | Limited | Vendor confirmed | AI Semantic Search needs 64-bit x86 and 8GB RAM. AI Semantic Index needs x74 or above; x90 not stated. AI Mode needs a 16GB VRAM GPU and none is listed for this model. |
| GPU expansion | Limited | Vendor confirmed | Low-profile cards only by card height (68.9 mm). Official list is RTX 3050 LP 6GB only, none flagged AI recommended. |
| Hardware acceleration | Limited | Vendor confirmed | No iGPU. Hardware transcoding only with a listed RTX 3050 LP. No AI accelerator category on the compatibility list. |
| AI photos | Yes | Vendor confirmed | QuMagie AI-assisted photo management. |
| AI video | Yes | Vendor confirmed | QVR Face listed under QTS only. |
| Surveillance AI | Yes | Vendor confirmed | QVR Surveillance 2 free channels plus 6 on registration; QVR Pro 8 free (QTS). |
Calculated from the TS-h2490FU's memory and each model's published size at each quantization. Rows marked Estimated are calculations, not benchmarks. Check any model.
| Model | Best quantization | Runs? | Memory needed | Expected performance | Verification |
|---|---|---|---|---|---|
| Llama 3.2 3BLlama, 3.21B | Q8 | Yes | 13GB | Not measured | Estimated |
| Llama 3.1 8BLlama, 8.03B | Q8 | Yes | 19GB | Not measured | Estimated |
| Qwen3 8BQwen, 8.19B | Q8 | Yes | 19GB | Not measured | Estimated |
| Gemma 3 12BGemma, 11.8B | Q8 | Yes | 24GB | Not measured | Estimated |
| Qwen3 14BQwen, 14.8B | Q8 | Yes | 28GB | Not measured | Estimated |
| Phi-4 14BPhi, 14.7B | Q8 | Yes | 28GB | Not measured | Estimated |
| Mistral Small 3 24BMistral, 23.6B | Q8 | Yes | 38GB | Not measured | Estimated |
| Gemma 3 27BGemma, 27B | Q8 | Yes | 44GB | Not measured | Estimated |
| Qwen3 30B A3BQwen, 30.5B | Q8 | Yes | 47GB | Not measured | Estimated |
| Qwen3 32BQwen, 32.8B | Q8 | Yes | 50GB | Not measured | Estimated |
| DeepSeek-R1 32B (Qwen distill)DeepSeek, 32.8B | Q8 | Yes | 50GB | Not measured | Estimated |
| Llama 3.3 70BLlama, 70.6B | Q8 | Yes | 98GB | Not measured | Estimated |
| Llama 4 Scout 17Bx16ELlama, 109B | Q8 | Yes | 149GB | Not measured | Estimated |
| Qwen2.5 72BQwen, 72.7B | Q8 | Yes | 101GB | Not measured | Estimated |
| Mistral Large 2 123BMistral, 123B | Q8 | Yes | 164GB | Not measured | Estimated |
One verdict per workload, from the TS-h2490FU's use-case scores and capability rows. Unknown means we cannot confirm it, not that it probably works.
The TS-h2490FU is a good fit for local LLM work, though it leans on the CPU. Ollama is vendor confirmed, local LLM support is limited, and the only listed GPU is the 6GB RTX 3050 LP, so inference is mostly CPU.
Private RAG is a good fit here, with caveats on the GPU side. RAG support is limited and vendor confirmed, with Open WebUI and Ollama both supported, but on-premises RAG needs a compatibility-tested GPU.
AI search is a good fit. Support is limited and vendor confirmed: Qsirch semantic search needs 64-bit x86 and 8GB of RAM, but AI Mode needs a 16GB VRAM GPU and none is listed for this system.
AI photos is a good fit. QuMagie AI-assisted photo management is listed as supported and vendor confirmed, and it does not depend on the limited GPU path.
Surveillance is a good fit. QVR Surveillance is vendor confirmed, with 2 free channels plus 6 on registration, and QVR Pro offers 8 free channels under QTS.
Home lab use is a good fit. Docker and virtual machines are both supported, and the system has a 16-core EPYC, 128GB of ECC memory as standard, and five PCIe Gen4 slots.
Business use is a good fit. It has ECC memory, dual 1100W power supplies, 4 x 25GbE on the 7302P models, 24 U.2 NVMe bays, a 5-year warranty and a 2U rackmount form.
Creative AI is a good fit, within a narrow GPU path. GPU expansion and hardware acceleration are both limited: only the 6GB RTX 3050 LP is listed, and hardware transcoding needs that card.
Plan on upgrading. 128GB is fitted as standard, while the TS-h2490FU accepts 4096GB of DDR4 ECC RDIMM, ECC included, which is the difference between 70B class models at higher-quality quantization, or large mixture-of-experts models at 4-bit and 70B class models at higher-quality quantization, or large mixture-of-experts models at 4-bit (Estimated).
Our estimate for the biggest model that fits in 4096GB is about 6813B parameters at 4-bit, after setting aside 8GB for the system. Long context windows reduce that figure (Estimated).
ECC support protects against silent memory errors on a box that runs around the clock.
Yes, through a 5x PCIe Gen4: Slot 1 x4, Slot 2 x8 (x4 when Slot 3 is used), Slot 3 x4, Slot 4 x8, Slot 5 x16 (x8 when Slot 4 is used). 25GbE adapters preinstalled in Slots 4 and 5 on 7302P models, Slot 5 on 7232P models. Cards up to 225 x 68.9 mm.. That is the single biggest upgrade for local AI on the TS-h2490FU: model layers move into fast graphics memory and replies speed up dramatically. Card size and power are the constraints (1100W supply).
As the storage half of a RAG setup it works well; as the model host it is limited. Small embedding models run fine on the TS-h2490FU, while the answering model needs more compute than it has.
For network access it has 4 x 25GbE + 2 x 2.5GbE. Fast enough that the network will not be the bottleneck.
| Configuration | RAM | SSD | Notes |
|---|---|---|---|
| TS-h2490FU-7232P-32G | 32GB | None | 2 x 25GbE. EPYC 7232P/7252 8C/16T. |
| TS-h2490FU-7302P-64G | 64GB | None | 4 x 25GbE. |
| TS-h2490FU-7232P-64G | 64GB | None | 2 x 25GbE. EPYC 7232P/7252 8C/16T. |
| TS-h2490FU-7302P-128G | 128GB | None | 4 x 25GbE. |
| TS-h2490FU-7302P-256G | 256GB | None | 4 x 25GbE. |
The closest QNAP model by bay count and AINAS Score. Only the specifications that differ are listed.
Best fit: home lab owners who want one box for storage, containers, media and a few AI services. That is where the TS-h2490FU scores highest (Home Lab 82/100), followed by Business (79/100).
Other AI Ready systems closest in overall score.
Workloads it suits
Yes, as an estimate. Qwen3 235B A22B needs 142GB at Q4 and 250GB at Q8, and both fit in memory on the TS-h2490FU (Estimated). The standard 128GB is below that, so you would need a larger memory configuration. The system accepts up to 4096GB of DDR4 ECC RDIMM. This is a calculation, not a benchmark, and no speed is implied.
Up to 4096GB. The memory is DDR4 ECC RDIMM and it is upgradeable. The standard model carries 128GB. QNAP's compatibility list tops out at 128GB modules, which gives 2TB across 16 slots, so reaching 4096GB needs 256GB modules that are not on that list.
Yes, but only a low-profile one. The GPU path is a PCIe x8 slot, and cards are limited to 68.9 mm in height. QNAP's official list has just the 6GB RTX 3050 LP, and none is flagged as AI recommended. Support is marked limited and vendor confirmed.
Two. QNAP lists 1100W x2 for the TS-h2490FU, with the full 1100W at 200-240Vac and 850W at 100-240Vac. That matters for rack planning, because the supplies feed a 16-core EPYC 7302P, up to 4096GB of DDR4 ECC RDIMM and 24 U.2 bays. The only listed graphics card, the 6GB RTX 3050 LP, is low-profile, so GPU draw is not the main load. Check which input voltage your rack provides before counting on the higher rating.
Yes. Container Station supports LXC/LXD and Docker. The system runs QuTS hero h6.1.0 or QTS 6.1.0. QNAP's Ollama guide for x86 QTS and QuTS hero systems is vendor confirmed, and Open WebUI is vendor confirmed with QTS 5.2 or QuTS hero h5.2 or later.
Five years. This is the standard QNAP warranty for the TS-h2490FU. It was released in 2020 and is not marked as end of life. The warranty applies across the configurations listed, from the 7232P 32GB model to the 7302P 256GB model.
Yes. Ollama support is documented by the manufacturer. See the capability table above for the source.
Yes. Open WebUI support is documented by the manufacturer. See the capability table above for the source.
Yes. AI photos support is documented by the manufacturer. See the capability table above for the source.
Yes. Surveillance AI support is documented by the manufacturer. See the capability table above for the source.
Partly. Hardware acceleration support is documented by the manufacturer. See the capability table above for the source.
It has 4 x 25GbE + 2 x 2.5GbE built in, and the 5x PCIe Gen4: Slot 1 x4, Slot 2 x8 (x4 when Slot 3 is used), Slot 3 x4, Slot 4 x8, Slot 5 x16 (x8 when Slot 4 is used). 25GbE adapters preinstalled in Slots 4 and 5 on 7302P models, Slot 5 on 7232P models. Cards up to 225 x 68.9 mm. can take a faster network card. The fastest port is 25GbE.