Two kinds of small desktop box now carry 128 GB of memory that the AI chip can use: NVIDIA's GB10 box, sold by ASUS as the Ascent GX10, and mini PCs built on the AMD Ryzen AI Max+ 395. Both can hold 100-billion-parameter models that no single consumer graphics card holds. This is the honest list of what runs, and how fast.
What decides the list
Three things set what a box can do. Memory size decides whether a model fits at all. Memory bandwidth, how fast the chip reads that memory, decides tok/s (tokens per second, roughly words per second). Both boxes are close: 273 GB/s for the GB10 and 256 GB/s for the Ryzen chip. Model type decides how much of that bandwidth each word costs. A dense model reads all its weights for every word. A MoE (mixture-of-experts) model reads only a few billion of them, so a 120B MoE can write faster than a 30B dense model.
Usable memory is a little less than the sticker. The selector counts 115 GB on the GX10. The Ryzen box shares one pool between the processor and the graphics side. AMD's guide says Linux limits the graphics side to about half of the RAM by default and shows how to raise it. AMD itself quotes up to 96 GB for graphics, the Windows setting; the selector assumes about 120 GB on Linux and 96 GB on Windows. Out of the box the GPU may see less than that, so a Ryzen box needs setting up.
Model by model
Q4 is a 4-bit model file, a trade of some quality for size. The numbers are for one user and a 32,000-token context (the amount of text the model holds in mind at once).
| Model | Type | Needs | GX10 | Ryzen AI Max+ 395 |
|---|---|---|---|---|
| gpt-oss-120b | MoE | about 67 GB | about 42 tok/s | about 40 tok/s |
| Qwen3.5-122B-A10B | MoE | about 81 GB | about 26 tok/s | about 25 tok/s |
| Nemotron 3 Super 120B-A12B | MoE | about 90 GB | about 21 tok/s | about 20 tok/s |
| GLM-4.5-Air | MoE | about 83 GB | about 14 tok/s | about 13 tok/s |
| Qwen3-Coder-Next 80B-A3B | MoE | about 52 GB | about 56 tok/s | about 54 tok/s |
| Qwen3.8-Flash-Next | MoE | about 83 GB (smallest file) | about 46 tok/s | about 44 tok/s |
| gpt-oss-20b | MoE | about 14 GB | about 54 tok/s | about 51 tok/s |
| Gemma 4 26B-A4B | MoE | about 20 GB | about 47 tok/s | about 45 tok/s |
| Qwen3.8 27B | Dense | about 22 GB | about 10 tok/s | about 10 tok/s |
| Llama 3.3 70B | Dense | about 56 GB | about 4 tok/s | about 4 tok/s |
Qwen3.8-Flash-Next shows the memory limit. Its smallest file, about 83 GB, fits. Its Q4 file, about 125 GB, does not fit on either box.
Read the table in three bands. MoE models run well: 40 tok/s or more is faster than most people read, and gpt-oss-120b, the headline reason to buy one of these boxes, sits there. The 120B-class models with more active weights, such as Nemotron 3 Super and GLM-4.5-Air, are slower but still usable for one person. Dense models of 27B to 31B land around 10 tok/s, which is readable but slow for a long answer. Dense 70B at about 4 tok/s is too slow for chat. The box holds it; it does not run it comfortably. If you want a large model, pick a MoE one.
A public anchor for gpt-oss-120b
The selector's 42 tok/s is a calculation. For a real-world check, the llama.cpp community's DGX Spark discussion logs runs of gpt-oss-120b on the same chip as the GX10, using llama-bench with the CUDA backend. These are public results, not EMD Lab measurements. The same model got faster between two software builds a few weeks apart.
| Build | Reads a prompt | Writes | Reads, 32K deep | Writes, 32K deep |
|---|---|---|---|---|
| 22 Oct 2025 | 1,737 | 45.9 | 1,232 | 34.3 |
| 2 Nov 2025 | 1,956 | 60.6 | 1,027 | 40.6 |
"Reads" is prompt processing, how fast the box digests your input. "Writes" is token generation, the tok/s you see. "32K deep" means 32,000 tokens are already in the context. Writing slows as the context fills, which matters for long documents. The Ryzen chip is close on writing speed in the selector, but it reads prompts more slowly: in the llama.cpp community results on a small 7B model it reads about 900 to 1,000 tok/s against about 3,100 to 3,700 for the DGX Spark. So the first word of an answer takes longer when you feed it a big file.
The differences that matter
- GX10: full CUDA, NVIDIA's software stack, so almost every AI tool runs without translation. It ships with NVIDIA DGX OS, a Linux distribution based on Ubuntu, and a ConnectX-7 network port rated at 200 Gb/s. Two units can be linked to share a larger model.
- Ryzen AI Max+ 395 box: runs Linux with ROCm (AMD's compute stack) or Vulkan, or Windows. Most chat tools work. Some newer tools and research code expect CUDA and need extra work here. Models and tools move fast, so check the one you need before you buy.
Who it suits
One to five people chatting with the same model at once is the realistic range. The selector counts how many 32K conversations fit in memory for gpt-oss-120b (about 40 on the GX10, about 44 on the Ryzen box), but memory is not the real limit: the speed is shared, so each extra user slows everyone down. Image and video models run on both boxes, but slowly, so treat them as an occasional job rather than a service.
If more people need the assistant, or you want fast answers at long context, move to a GPU server. The QNAP AI server is one route, and the AI workstation with two RTX PRO 6000 Max-Q cards is another. Our hardware guide compares the options side by side.
The short version
- A 128 GB box holds 100B-class MoE models, with gpt-oss-120b at an estimated 40 to 42 tok/s.
- Dense models of 27B to 31B run near 10 tok/s; dense 70B runs near 4 tok/s.
- GX10: CUDA, DGX OS, a 200 Gb/s port to link two units. Ryzen AI Max+ 395: Linux or Windows, needs setting up, reads long prompts more slowly.
- Plan it for one to five people. Beyond that, a GPU server is the better tool.
All speeds in the first table are calculated estimates. Try your own model in the model finder.