96GB max RAM · 5 bays · 10GbE
AINAS Score 65
$919.00
Amazon US · Verified
Ships from the US; delivery and import charges shown at checkout

AI ReadyQNAPDiscontinued
Why AI Ready: Xeon E-2236 with PCIe Gen 3 expansion and a 128GB memory ceiling meets the Xeon ready rule; no integrated GPU is listed, the best listed GPU is the 5GB Quadro P2000, and QNAP documents no GPU inference on this model.
Updated
Can the QNAP TS-h1683XU-RP 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-h1683XU-RP: Intel Xeon E-2236 (E-2136 on E2136-128G) does the work with up to 128GB 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-h1683XU-RP suits a business or home lab buyer who wants a 16-bay rackmount NAS that can also host local AI in containers. It has a 6-core Intel Xeon E-2236 and 128GB of DDR4 ECC memory. It runs QTS 6.1.0 or QuTS hero h6.1.0, with Docker and virtual machines supported. Four 10GbE ports and four 1GbE ports cover networking, and a PCIe Gen 3 x8 slot leaves a path for a GPU. Two 500W power supplies are fitted and the warranty is 3 years. Model memory is plentiful: Mistral Large 2 123B fits at Q4 (Estimated), and the AINAS Score is 66.
The main catch is that QNAP has discontinued the TS-h1683XU-RP, and both SKUs are marked end of life, which matters to anyone planning years of use. The 128GB of memory is also fixed, with no expansion, so today's model list is the ceiling: Mistral Large 2 123B at Q4 and Llama 3.3 70B at Q8 fit (Estimated), and Qwen3 235B A22B does not. GPU choice is narrow too. The listed cards top out at 5GB of VRAM, and slot 2, the only x8 slot, holds a preinstalled 10GbE adapter. Anyone planning GPU-heavy inference should look for a current model with a free x8 or x16 slot, a GPU list with far more VRAM, and a memory ceiling above 128GB.
No verified price
We have not verified a listing for the QNAP TS-h1683XU-RP 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 applies (x86_64, 16GB or more for 7B models, GPU optional and must be in Container Station mode). 128GB allows CPU inference; the largest listed GPU is the 5GB Quadro P2000. |
| Ollama | Yes | Vendor confirmed | QNAP Container Station Ollama tutorial for QTS and QuTS hero on x86_64. |
| Open WebUI | Yes | Vendor confirmed | QNAP Open WebUI tutorial, QTS 5.2 or QuTS hero h5.2 or later. |
| RAG | Limited | Unverified | Qsirch lists on-premises RAG with no model named. Not confirmed for this model. |
| AI search | Yes | Vendor confirmed | Qsirch AI Semantic Search: 64-bit x86, 5.0.1 or later, 8GB RAM. Qsirch says the AI Semantic Index and AI Mode need an x74 series NAS or later; not confirmed for this model, and AI Mode also needs a 16GB or larger VRAM GPU. |
| GPU expansion | Limited | Vendor confirmed | Slot 2 (Gen 3 x8) holds a preinstalled 10GbE adapter; 220 mm card limit. GPU list tops out at Quadro P2000 5G and GTX 1650 LP. |
| Hardware acceleration | Limited | Vendor confirmed | No integrated graphics listed. Listed GPUs support Hardware Transcoding. No AI accelerator list for this model. |
| AI photos | Yes | Vendor confirmed | QuMagie listed under QTS only. |
| 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 channels (QTS only). |
Calculated from the TS-h1683XU-RP'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 | Q4 | Yes | 89GB | Not measured | Estimated |
| Qwen2.5 72BQwen, 72.7B | Q8 | Yes | 101GB | Not measured | Estimated |
| Mistral Large 2 123BMistral, 123B | Q5 | Yes | 112GB | Not measured | Estimated |
One verdict per workload, from the TS-h1683XU-RP's use-case scores and capability rows. Unknown means we cannot confirm it, not that it probably works.
The TS-h1683XU-RP is a good fit for local LLM work. Ollama is vendor confirmed through Container Station, and 128GB of memory allows CPU inference, though vendor support for local LLM is marked limited.
Private RAG is an OK fit, not a strong one. Qsirch lists on-premises RAG, but that row is Unverified and not confirmed for this model, while Open WebUI is vendor confirmed.
AI search is a good fit. Qsirch AI Semantic Search is vendor confirmed, though its AI Semantic Index and AI Mode are not confirmed for this model and AI Mode needs a 16GB or larger VRAM GPU.
AI photos is a good fit. QuMagie is vendor confirmed on this system, but it is listed under QTS only, so QuTS hero users should check before relying on it.
Surveillance is a good fit. QVR Surveillance gives 2 free channels plus 6 on registration, and QVR Pro gives 8 free channels, both under QTS only.
Home lab use is a good fit. Docker and virtual machines are both supported, and 16 bays, 128GB of ECC memory and four 10GbE ports give plenty of room to experiment.
Business use is a good fit. ECC memory, two 500W power supplies, a 3 year warranty and a rackmount chassis suit server rooms, though the system is discontinued.
Creative AI is a good fit on paper. The PCIe Gen 3 x8 slot accepts a GPU, but GPU expansion is limited, since that slot holds a preinstalled 10GbE adapter and the listed GPUs top out at 5GB.
128GB of DDR4 ECC UDIMM is all the TS-h1683XU-RP will ever have; it is not user-upgradeable. Every model you run has to fit inside that figure alongside the operating system.
Our estimate for the biggest model that fits in 128GB is about 200B parameters at 4-bit, after setting aside 8GB for the system. Long context windows reduce that figure (Estimated).
ECC memory support is a plus for always-on systems that also hold important files.
Yes, through a 4x PCIe: slot 1 Gen 2 x4, slot 2 Gen 3 x8, slots 3 and 4 Gen 3 x4; slots 2 and 4 preinstalled with 10GbE adapters; cards up to 220 x 78.9 x 18.76 mm. That is the single biggest upgrade for local AI on the TS-h1683XU-RP: model layers move into fast graphics memory and replies speed up dramatically. Card size and power are the constraints (500W 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-h1683XU-RP, while the answering model needs more compute than it has.
The manufacturer also provides AI search across your files (documented by the manufacturer), which covers simple document questions without setting up your own stack.
For network access it has 4 x 10GbE + 4 x 1GbE. Fast enough that the network will not be the bottleneck.
| Configuration | RAM | SSD | Notes |
|---|---|---|---|
| TS-h1683XU-RP-E2236-128G | 128GB | None | Xeon E-2236 6C/12T, 128GB ECC. Marked EOL. |
| TS-h1683XU-RP-E2136-128G | 128GB | None | Xeon E-2136 6C/12T, 128GB ECC. Marked EOL. |
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-h1683XU-RP scores highest (Home Lab 75/100), followed by Business (73/100).
Less suited to buyers who want a model with a long support life ahead.
Other AI Ready systems closest in overall score.
96GB max RAM · 5 bays · 10GbE
AINAS Score 65
$919.00
Amazon US · Verified
Ships from the US; delivery and import charges shown at checkout
96GB max RAM · 5 bays · 10GbE
AINAS Score 67
$519.00
Amazon US · Verified
Ships from the US; delivery and import charges shown at checkout
192GB max RAM · 12 bays · 2 x 10GbE
AINAS Score 64
No verified price
Workloads it suits
Yes, at Q4 only. Mistral Large 2 123B needs 73GB at Q4 and fits in the 128GB of memory (Estimated). At Q8 it needs 130GB and does not fit (Estimated). This is a memory calculation, not a benchmark. Without a GPU, inference runs on the CPU. Qwen3 235B A22B needs 142GB at Q4 and does not fit (Estimated).
No. The system ships with 128GB of DDR4 ECC UDIMM memory and the maximum is also 128GB, so there is no headroom to add. Buyers who expect to need more memory later should look for a system with a higher memory ceiling and free memory slots.
Yes. The 128GB it ships with is DDR4 ECC UDIMM, and the Xeon E-2236 platform supports it. ECC helps with reliability on systems that run continuously, such as servers that host virtual machines or containers for business use.
Yes, with limits. It has a PCIe Gen 3 x8 slot, but slot 2 holds a preinstalled 10GbE adapter and cards are limited to 220mm. The listed GPUs top out at the Quadro P2000 5G and GTX 1650 LP. QNAP documents no GPU inference on this model.
Yes at Q4, no at Q8 (Estimated). Llama 4 Scout 17Bx16E needs 67GB at Q4, which fits in the fixed 128GB, while Q8 needs 117GB and does not fit (Estimated). The listed GPUs stop at the 5GB Quadro P2000, so the model would run on the 6-core Xeon E-2236 and system memory. Ollama is vendor confirmed through Container Station on x86_64, and the 128GB leaves room for containers and virtual machines beside the model. These are memory calculations, not benchmarks.
Two configurations exist, both with 128GB of ECC memory. One uses the Xeon E-2236 and the other the Xeon E-2136, both with 6 cores and 12 threads. Both are marked end of life, and both share the same 16 bays and four 10GbE ports.
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 10GbE + 4 x 1GbE built in, and the 4x PCIe: slot 1 Gen 2 x4, slot 2 Gen 3 x8, slots 3 and 4 Gen 3 x4; slots 2 and 4 preinstalled with 10GbE adapters; cards up to 220 x 78.9 x 18.76 mm can take a faster network card. The fastest port is 4 x 10GbE.
Every field records its source. Specifications were last checked on . Spot an error? Send a correction with a manufacturer link.