The ASUS Ascent GX10 Showed Up. Here's What I'm Actually Running on It.

I posted "tomorrow is the day" like a kid waiting on a parcel. Then the ASUS Ascent GX10 actually arrived, and I did what anyone with a cube that expensive should do: I took pictures, I turned the sound up on a silly soundtrack, and I started putting real work on it instead of quoting the press deck.

This is not a lab review with charts yet. This is what I told X (@wikiwayne) in public, expanded into the notes I wish existed the week the box landed.

What I actually bought

The GX10 is a standalone desktop AI computer, not a card. NVIDIA's name for the family is DGX Spark. ASUS's name for this chassis is Ascent GX10. Inside is the NVIDIA GB10 Grace Blackwell Superchip: a 20-core Arm CPU (10× Cortex-X925 + 10× Cortex-A725) glued to a Blackwell GPU, sharing 128GB of LPDDR5x unified memory.

The numbers that matter on my desk:

• 128GB coherent unified memory — this is the whole pitch. A 24GB gaming GPU is a different animal.
• About 1 petaFLOP of FP4 AI on paper. Treat that as a marketing ceiling until I publish my own tokens/sec.
• NVIDIA DGX OS (Ubuntu-based) preloaded. You are not installing Windows on this and pretending it is a Steam Machine.
• ConnectX-7 plus 10GbE if you ever stack two of them.
• 150 × 150 × 51mm, about 1.5kg, 240W USB-C brick. It looks like a network appliance that wandered into the wrong product photo.

If you are still shopping GPUs for a regular PC, start with the best GPU for local AI (2026) and the VRAM requirements guide. The GX10 is a different purchase: you are buying a tiny DGX, not a 5090.

Why I wanted it

I write about running open-weight models locally. At some point the honest move is to stop renting context windows and put a box on the bench that can hold a serious model without swapping itself to death.

The Spark-class machines exist for that: local agents, local coding, local image work, no prompt leaving the room. ASUS even markets the GX10 as agent-ready (OpenClaw, NemoClaw, Hermes). I care less about the brand names and more about whether I can keep DeepSeek, Qwen, and a coding agent on metal I own.

A buddy joked that mine was "really an open-box return." Buys beware, sure — inspect the unit, keep the serial, don't skip the unbox video. Mine powered on. The rest is workload.

First week on the bench

I did not start with a 70B bake-off. I started with work I already owed myself:

1. Local agent + DeepSeek V4 Flash to ship a real thing: Moon Trail. I wrote that up separately because it is the first project that used the box instead of posing next to it.
2. Qwen 3.8 27B — I posted "we believe in you" because 27B is the size that used to feel like a stretch on a 24GB card and now fits in the "of course it fits" bucket on 128GB unified.
3. Photos. The cube photographs like a speaker. Stack two and you get the leaning-tower joke I posted next to Pisa. That is not a benchmark. That is me being glad it exists.

I have not published a tokens-per-second table yet. When I do, it will be measured on this unit with named quants and named runners — not a screenshot from a keynote.

What 128GB unified memory changes

On a 12–24GB NVIDIA card you live in quantization math. Q4 vs Q8, context vs KV cache, "can I load 14B if I offload six layers." That math still matters — see quantization explained and Q4 vs Q8 — but the GX10 moves the ceiling.

With 128GB coherent memory you can:

• Keep a 27B–70B-class chat model resident without treating it like a science fair.
• Run an agent loop (tools, files, a second helper model) without immediately paging to disk.
• Fine-tune or experiment with larger checkpoints than a consumer GPU will even map.

You still pick a runner. Ollama vs LM Studio, llama.cpp, and whatever NVIDIA ships on DGX OS are different doors into the same memory pool. I am using what the agent stack already speaks (Hermes + a local DeepSeek Flash build) and I will add a cleaner Ollama path as I settle the box in.

What I am not claiming

• I am not claiming the GX10 beats a rack DGX or a cloud H100 on every job. It is a desk machine.
• I am not claiming FP4 petaFLOPs until I show a workload.
• I am not claiming you should throw away a perfectly good 3090/4090 if your models already fit. Buy this when unified memory and a quiet local agent box are the bottleneck, not when you need rasterization.

Who this box is for

Buy a Spark-class cube if you:

• Run local LLMs as a daily driver, not a demo.
• Want agents that can see your repo without shipping the repo to a vendor.
• Care about the privacy tradeoff.

Skip it if you:

• Only need 8B chat. A used GPU from the budget guide is cheaper and simpler.
• Need Windows, CUDA-as-you-know-it on x86, or a card that also plays games.
• Want to swap VRAM later. You cannot.

What's next

Moon Trail is already live. Next on this hardware: a proper model card for whatever Qwen 3.8 27B build I keep, a Flash vs Pro note for DeepSeek V4, and an honest tokens/sec pass I can defend. Until then, the local AI model tracker stays the living index.

If you are on the same Spark/GX10 wave, tell me what you loaded first — X (@wikiwayne). I will keep posting the unglamorous version.

Related: nvidia vs amd gpu local llm 2026