64GB max RAM · 6 x M.2 · 10GbE
AINAS Score 35
£979
Amazon UK · Verified

AI CapableTerraMaster
Why AI Capable: Intel Celeron N5095 (4 cores, UHD Graphics, no NPU), 4GB shipped and 32GB max DDR4, no PCIe or other GPU path: Docker-capable x86 but only tiny models.
Updated
Can the TerraMaster U4-423 run local AI?
The TerraMaster U4-423 can run small local models, with limits. Its Intel Celeron N5095 and 32GB memory ceiling suit models up to the 14B class, with 32B class models a tight fit (Estimated), so treat it as a storage-first NAS that can host light AI tools.
There is no way to add a discrete GPU, so all inference runs on the CPU and basic integrated graphics.
Local LLM support is not yet verified.
The U4-423 suits a buyer who wants a 4-bay rackmount NAS for file storage, Docker containers and camera recording, and is happy to keep AI as a light extra. It has an Intel Celeron N5095 with 4 cores, two 2.5GbE ports, one M.2 slot and Docker and virtual machine support on TOS 7. Memory starts at 4GB and takes up to 32GB of DDR4 non-ECC SODIMM, so you can add RAM yourself. Surveillance Manager handles up to 32 ONVIF IP cameras, and the Intel iGPU does video transcoding up to 4K60. The AINAS Score is 35, which reflects an AI class of capable with firm limits.
The main thing that rules the U4-423 out is the lack of any GPU path. It has no PCIe slot, no USB4, no Thunderbolt and no OCuLink, so the integrated Intel UHD Graphics is all you get, and it only handles video. There is also no NPU. Anyone planning to run larger local models at usable responsiveness, or to add a graphics card later, should look for a system with a PCIe slot or an OCuLink or Thunderbolt port, plus a higher memory ceiling and an NPU or a GPU path. Every AI workload on this page carries a poor verdict.
No verified price
We have not verified a listing for the TerraMaster U4-423 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 | Unverified | Low-power CPU with small shipped RAM and no GPU path; only very small models via Docker would be practical. TerraMaster blog describes Ollama/Open WebUI via Docker on x86 TOS NAS generally, not this model. |
| Ollama | Limited | Vendor confirmed | Runs as a Docker container via TOS Docker Manager per TerraMaster blog (x86 TOS NAS generally). Practical only for tiny models on this hardware. |
| Open WebUI | Yes | Vendor confirmed | Open WebUI container via Docker; same TerraMaster blog source. |
| RAG | Limited | Unverified | Possible through Open WebUI or similar containers; no TerraMaster or community source showing RAG on TOS. |
| AI search | Limited | Vendor confirmed | TOS 7 OpenClaw agent and AI tagging allow natural-language queries; setup uses an API key from a provider such as OpenAI, so cloud LLM. Global search is an inverted index, not semantic. |
| GPU expansion | No | Vendor confirmed | Spec lists no PCIe slot; no USB4, Thunderbolt or OCuLink. |
| Hardware acceleration | Yes | Vendor confirmed | Hardware transcoding engine H.264/H.265/MPEG-4/VC-1 up to 4K60 via Intel iGPU. Video only, no NPU. |
| AI photos | Yes | Vendor confirmed | TOS 7 Photos: AI face recognition and AI scene recognition (80+ categories). |
| AI video | No | Unverified | No AI video feature found in TOS 7 material. |
| Surveillance AI | Yes | Vendor confirmed | TOS 7 Surveillance Manager supports up to 32 ONVIF IP cameras. |
Calculated from the U4-423'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 | Q4 | Yes | 25GB | Not measured | Estimated |
| Gemma 3 27BGemma, 27B | Q4 | Yes | 29GB | Not measured | Estimated |
| Qwen3 30B A3BQwen, 30.5B | Q4 | Yes | 31GB | Not measured | Estimated |
| Qwen3 32BQwen, 32.8B | Q4 | Yes | 32GB | Not measured | Estimated |
| DeepSeek-R1 32B (Qwen distill)DeepSeek, 32.8B | Q4 | Yes | 32GB | Not measured | Estimated |
| Llama 3.3 70BLlama, 70.6B | None fits | No | 60GB at Q4 | Not measured | Estimated |
| Llama 4 Scout 17Bx16ELlama, 109B | None fits | No | 89GB at Q4 | Not measured | Estimated |
| Qwen2.5 72BQwen, 72.7B | None fits | No | 65GB at Q4 | Not measured | Estimated |
| Mistral Large 2 123BMistral, 123B | None fits | No | 96GB at Q4 | Not measured | Estimated |
One verdict per workload, from the U4-423's use-case scores and capability rows. Unknown means we cannot confirm it, not that it probably works.
Local LLM work is a poor fit on the U4-423. The capability is limited and Unverified, there is no GPU path, and the Celeron CPU ships with 4GB of RAM, so only very small models via Docker would be practical.
Private RAG is a poor fit here. RAG is limited and Unverified, with no TerraMaster or community source showing it on TOS, though Open WebUI is vendor confirmed as a Docker container.
AI search is a poor fit. TOS 7 offers natural-language queries as a limited, vendor-confirmed feature, but it needs an API key from a cloud provider, and the built-in global search is not semantic.
AI photos are rated poor despite the feature being present. TOS 7 Photos has AI face recognition and scene recognition across 80+ categories (vendor confirmed), but the CPU has no NPU or GPU path to speed it up.
Surveillance is rated poor for AI use. Surveillance Manager is vendor confirmed for up to 32 ONVIF IP cameras, but the record shows no AI video feature in TOS 7 material and no NPU.
Home lab use is rated poor overall. Docker and virtual machines are both supported on the x86 CPU, but the 4GB of shipped memory and the absence of any PCIe expansion limit what you can run.
Business use is rated poor. The U4-423 has no ECC memory, its RAM is non-ECC SODIMM, and the 4GB fitted is small for business services, which weakens it for dependable deployment.
Creative AI is a poor fit. GPU expansion is vendor confirmed as not supported, and the Intel iGPU accelerates video transcoding only, so image or video generation has no hardware path.
The U4-423 ships with 4GB of DDR4 non-ECC SODIMM and supports up to 32GB. Upgrading to the maximum lifts the model ceiling from only very small models of around 1B to 3B parameters to models up to the 14B class, with 32B class models a tight fit (both Estimated).
As a rule of thumb, a 4-bit model needs about 0.6GB per billion parameters plus room for context. After leaving around 8GB for the NAS operating system, 32GB leaves space for a model of roughly 40B parameters at most (Estimated).
No. The U4-423 has no PCIe slot or OCuLink port for a graphics card. Inference runs on the CPU, with Intel UHD Graphics handling media tasks, so keep to small models.
With limits. The U4-423 can host a small RAG stack such as AnythingLLM or Open WebUI with a compact embedding model, but memory and compute restrict how large the answering model can be. It suits a personal document library more than a team knowledge base.
Networking is 2 x 2.5GbE. Fine for one or two users; indexing a large archive over the network will be slower.
| Configuration | RAM | SSD | Notes |
|---|---|---|---|
| 4GB | 4GB | None | Only configuration; 1x 4GB DDR4, 2 slots. |
The closest TerraMaster model by bay count and AINAS Score. Only the specifications that differ are listed.
The U4-423 scores best for Home Lab (37/100) and Business (37/100). It suits home lab owners who want one box for storage, containers, media and a few AI services.
It is a weaker fit for buyers who expect to add a graphics card later and for teams moving large files over the network every day.
Other AI Capable systems closest in overall score.
Estimated: yes at Q4, no at Q8. DeepSeek-R1 32B (Qwen distill) needs 20GB at Q4 and 35GB at Q8, and the U4-423 takes up to 32GB. That is a memory calculation, not a benchmark. The Celeron N5095 has no GPU path or NPU, and the vendor says only tiny models are practical, so memory fit does not mean usable responsiveness. Llama 3.3 70B does not fit at Q4 (Estimated).
Yes. It ships with 4GB and takes up to 32GB of DDR4 non-ECC SODIMM. The shipped configuration is a single 4GB module in one of 2 slots, so there is a free slot for expansion. The memory sits in slots rather than being soldered. Because it is non-ECC, a larger kit adds capacity but no error correction.
No. The U4-423 uses DDR4 non-ECC SODIMM, and ECC is listed as not supported. Memory is upgradeable up to 32GB, but only with non-ECC modules. Buyers who need error-correcting memory for data integrity should look for a system that lists ECC RAM in its specification.
No. The U4-423 has no PCIe slot, no USB4, no Thunderbolt and no OCuLink, and GPU expansion is vendor confirmed as not supported. The only graphics hardware is the integrated Intel UHD Graphics, which handles video transcoding up to 4K60 for H.264, H.265, MPEG-4 and VC-1.
Yes. The U4-423 has 1 M.2 slot alongside its 4 drive bays, which suits an NVMe cache or a small fast volume for Docker containers. With no PCIe slot, that single M.2 slot is the only internal option for flash storage, and the two 2.5GbE ports carry all network traffic. Memory is a separate limit, rising from the 4GB fitted to 32GB of DDR4 non-ECC SODIMM.
Yes to both. Docker is listed with TOS Docker Manager, and virtual machines are supported through TOS 7 KVM on x86. The 4GB of shipped memory is tight for either, so adding RAM up to the 32GB maximum makes sense if you plan to run several containers or a VM.
Partly. 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.
Yes. Hardware acceleration support is documented by the manufacturer. See the capability table above for the source.
It has 2 x 2.5GbE built in. The fastest port is 2 x 2.5GbE.
Every field records its source. Specifications were last checked on . Spot an error? Send a correction with a manufacturer link.