Good fit
You want to run quantized 70B models, CUDA-oriented tools, containers or memory-heavy local inference from a very small desktop system.
Strength: memory plus NVIDIA stackCompact NVIDIA AI computer
ASUS Ascent GX10 for local LLMs: buyer's guide
The Ascent GX10 combines NVIDIA's GB10 Grace Blackwell, 128GB of coherent unified memory and DGX OS in a six-inch-wide system. It is compelling for large local models and the NVIDIA software stack, but only when Arm, Linux and limited internal upgrades fit your workflow.
Quick verdict
Who should buy the ASUS Ascent GX10?
The GX10 is not an ordinary Windows PC with a fast NPU. It is a compact Linux development machine for local AI, built around a large shared memory pool and NVIDIA tooling.
Check first
Every essential application must support Arm64 and DGX OS. A Linux version alone does not guarantee Arm compatibility.
Platform: Arm64 and LinuxPoorer fit
You need a general gaming PC, Windows-only software, replaceable memory, easy SSD upgrades or maximum speed for models below 32GB.
Alternative: RTX 5090 towerSpecs that matter to a buyer
What is inside the Ascent GX10?
ASUS lists 1TB, 2TB and 4TB versions. Always verify the model number and SSD capacity in the offer rather than relying on the shared product name.
| Processor | NVIDIA GB10 Grace Blackwell with a 20-core Arm CPU and integrated Blackwell GPU |
|---|---|
| Model memory | 128GB LPDDR5X coherent unified memory, up to 273 GB/s |
| Storage | 1TB PCIe 4.0, 2TB PCIe 4.0 or 4TB PCIe 5.0; one M.2 2242 slot |
| Operating system | NVIDIA DGX OS with the NVIDIA AI software stack |
| Networking | 10GbE, Wi-Fi 7, Bluetooth 5.4 and ConnectX-7 for a second node |
| Ports | four USB-C ports, HDMI 2.1a and Kensington lock slot |
| Size and power | 5.91 × 5.91 × 2.01 in; 240W adapter and up to 180W device input |
ASUS's data sheet says the SSD is not user-changeable and opening the chassis may void the warranty. Choose enough storage before purchase.
Plan model size realistically
Which local LLMs fit in 128GB?
The CPU, GPU, operating system and runtime share the 128GB pool. This table is capacity planning with headroom, not a speed guarantee.
| Model class | Capacity outlook | What to check |
|---|---|---|
| 7B to 32B, Q4/Q5 | ample headroom | Desired throughput matters more than raw capacity in this range. |
| 65B to 70B, Q4 | practical fit | Context length, KV cache, concurrent sessions and extra services still consume memory. |
| 100B to 120B, heavily quantized | configuration dependent | Check file size and runtime before downloading; long context can consume the remaining headroom quickly. |
| up to 200B in platform claims | not a blanket promise | NVIDIA's figure depends on precision, architecture and workload and does not define usable context or response speed. |
An RTX 5090 tower is generally the speed-focused route for models that fit fully into 32GB. The GX10 instead emphasizes capacity and an integrated NVIDIA development stack.
Before ordering
Six checks that materially affect the purchase
Arm64 compatibility
Verify every essential runtime, Python dependency, extension and container architecture. A product offering Linux support may still be x86-only.
Choose storage up front
Large models, multiple quantizations and RAG data can fill 1TB quickly. A 4TB version makes sense for a substantial local model library.
Size model and context
Do not compare parameter counts alone. Model weights, KV cache, runtime and concurrent services must fit together.
Define its network role
10GbE is useful when several devices call the local model server or you regularly move large datasets.
Separate appliance and desktop roles
Treat the GX10 as an AI appliance. A conventional tower is more flexible for gaming, Windows-only applications and replaceable parts.
Warranty and seller
Confirm model number, SSD, package contents, warranty region, return terms and seller. Similar-looking listings can differ in storage and support.
Software before hardware price
Ollama, containers and LM Studio on the GX10
Ollama's current hardware list explicitly includes GB10 systems. LM Studio also supports Linux on Arm64 in principle; still match the current app release to DGX OS and your model formats before buying. For NVIDIA containers, PyTorch and TensorRT-LLM, the preconfigured stack is the main advantage over a conventional mini PC.
The main alternatives
GX10, Ryzen AI Max+ 395 or RTX 5090?
| Hardware route | Greatest strength | Main limitation | Best suited to |
|---|---|---|---|
| ASUS Ascent GX10 | 128GB plus NVIDIA software stack | Arm/Linux and few internal upgrades | CUDA-oriented development and large local models |
| Ryzen AI Max+ 395, 128GB | x86 PC with a large shared pool | check Radeon backend and UMA configuration | Windows/Linux desktop use and 70B capacity |
| GeForce RTX 5090, 32GB | high bandwidth and broad CUDA support | 70B usually needs slower offload | fast 8B to 32B inference |
Verify the exact version
ASUS Ascent GX10 with 128GB and 4TB
Confirm model number, SSD capacity, DGX OS, warranty region and seller. Do not choose from the shared product title alone.
* Paid link. We may earn a commission if you buy; your price is unchanged. As an Amazon Associate we earn from qualifying purchases. Verify configuration, seller, price and availability on the destination page.
Common questions
Using an ASUS Ascent GX10
Can the GX10 run a 70B LLM locally?
Its capacity is sufficient for many quantized 70B models with headroom. Actual usability also depends on format, quantization, context, runtime and your speed expectations.
Is the Ascent GX10 a Windows PC?
No. It uses an Arm processor and Linux-based NVIDIA DGX OS. Choose it only when your workflow supports that platform.
Can I upgrade the memory or SSD?
The 128GB memory is integrated. ASUS's data sheet also says the SSD is not user-changeable and opening the chassis may affect the warranty.
Is the GX10 faster than an RTX 5090?
Not universally. The RTX 5090 is the speed-focused choice for models that fit fully in 32GB. The GX10 provides much more shared capacity and a different software and platform focus.
Technical primary sources
Check the specifications yourself
Hardware data comes from the ASUS data sheet and NVIDIA's DGX Spark hardware guide. Manufacturer model-size claims are not independent performance benchmarks.