ASUS ProArt GR1X: Powerful Enough for AI, But Is It Powerful Enough for Multiple Users?

For a handful of people, probably yes; for a department, probably not. The ASUS ProArt GR1X has enough memory to hold a large model, but serving several users at once depends on three other things: how fast the chip can generate tokens for everyone combined, how much memory each user's context takes, and whether the serving software on Windows can batch requests efficiently. On the closest published reference, a Linux DGX Spark with a similar memory system, gpt-oss-120b dropped from about 55 tokens per second for one user to about 12 per user at 16 users. The GR1X runs Windows, where the strongest multi-user engines don't run natively, so expect that as an upper bound.

How does speed change as more users join?

Source: llama.cpp batched benchmark of gpt-oss-120b on an NVIDIA DGX Spark (GB10, Linux), 4,096-token prompt and 32-token reply, as published by eurokommerz (23 Sep 2026, updated 5 Nov 2026) citing the llama.cpp maintainers' February 2026 build. Combined output is per-user speed × users (our arithmetic). This is not a GR1X measurement. No multi-user benchmark of an RTX Spark Windows system had been published at the time of writing.

The pattern is what matters. Adding users raises the machine's total output, because one pass through the model's weights can serve several requests at once. But each person waits longer for their own reply. Around 20 tokens per second still feels interactive for chat. Below about 10 it feels slow, though background agents and batch jobs can live with it. A community-patched vLLM build on the DGX Spark reached about 173 tokens per second combined at 10 users and 299.5 at 30. That needed six custom patches on Linux, and it shows what better batching software can add.

Why does every extra user need more memory?

Each conversation keeps its own context in a KV cache, and that memory adds up per user. Assuming about 4 GB for the engine and overheads, here's roughly how many 32K-token conversations fit alongside the model (our arithmetic, BF16 KV cache unless stated):

• Qwen3-32B at 4-bit (19.8 GB) in a 64 GB GPU share: about 4–5 conversations, or about 9 with an 8-bit KV cache.
• Llama 3.3 70B at 4-bit (42.5 GB) in a 96 GB GPU share: about 4–5 conversations, or about 9 with an 8-bit cache.
• gpt-oss-120b (65.3 GB) in a 96 GB GPU share: about 22 conversations at 32K, or about 5 at 128K, because its cache is unusually small.
• Llama 3.1 8B at 4-bit (4.9 GB) in a 64 GB share: about 12–13 conversations.

Two caveats. On 128 GB RTX Spark laptops, Windows gives the GPU about 64 GB by default; ASUS's ProArt P16 lets you raise that to about 96 GB, but the GR1X's setting wasn't documented at the time of writing. And these counts assume every user fills their context. Real chats are often shorter, while long document work can be much longer. Our memory sizing guide has the per-model context figures.

A GR1X can hold a big model for one person very comfortably. Serving a team means multiplying the context memory and dividing the speed, and that's where small machines run out of room.

What does Windows change for serving?

Quite a lot, as of November 2026:

• The best batching engines aren't native. vLLM has no native Windows support, and TensorRT-LLM's installation is Linux-only. vLLM can run in WSL2, which XDA found works with GPU access on an RTX Spark laptop, or through an unofficial native port. Ollama and llama.cpp run natively.
• Defaults are set for one person. Ollama's default context is 4,096 tokens unless you change it, so a shared setup needs tuning before it handles long documents.
• Memory is split in advance. Unlike Linux on the DGX Spark, Windows gives the GPU a fixed share, so a multi-user setup has to be sized ahead of time.
• Security and updates. Smart App Control can block unsigned AI tools (XDA). Windows Home and Pro have historically limited how long updates can be paused, so plan reboot windows for a machine others rely on.

None of this stops a small team. It does mean a GR1X serving several people is more of a do-it-yourself setup than a Linux AI box, where vLLM and SGLang are standard.

How many people can realistically share one?

For interactive chat with a mid-size or mixture-of-experts model, a handful of people at once is realistic. Usage is bursty, so a team of perhaps 5–10 people who don't all type at the same moment may share one machine comfortably (our rough estimate, not a measured figure). For agents that run in the background, more parallel tasks are possible because nobody is watching each token. For a 70B dense model, multi-user is hard: at about 7 tokens per second as a theoretical single-user ceiling on about 300 GB/s of bandwidth, splitting that between users gets slow quickly. Our 70B on the GR1X post covers this.

If the GR1X is mainly running agents for its owner, see running AI agents 24/7. If you're weighing two units to serve more people, see two GR1X units vs one GPU workstation.

When should a team look at something bigger?

• A Linux GB10 box such as the ASUS Ascent GX10 or DGX Spark runs vLLM and SGLang natively, lets the GPU use most of its 128 GB, and two units can be linked over ConnectX-7. Our GB10 for 20 concurrent users post covers that route.
• An RTX PRO 6000 workstation has 96 GB at 1,792 GB/s, about six times the GR1X's bandwidth, and can be split into isolated instances with MIG.
• A server or DGX Station is the step for whole departments. See building a corporate ChatGPT on GB10 for the in-between options.

Buying in Singapore

The ASUS ProArt GR1X is listed as price on request. ASUS hasn't published a price. As a SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price, expect roughly S$9,000–12,000 for the 128 GB models a shared setup would need, including GST. For team serving on Linux, the ASUS Ascent GX10 128 GB / 1 TB is S$10,350 and the NVIDIA DGX Spark 4 TB is S$9,300; the RTX PRO 6000 Blackwell Workstation Edition card is S$29,100 (all listed 8 Nov 2026). Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

How many users can one ASUS ProArt GR1X support? No GR1X multi-user test had been published at the time of writing. Based on a similar Linux GB10 system and our memory arithmetic, a handful of simultaneous chat users is realistic with mid-size or mixture-of-experts models; a whole department is not.

Can I run vLLM on the GR1X? Not natively on Windows. vLLM can run inside WSL2, and an unofficial native Windows on Arm port exists. Ollama and llama.cpp run natively.

Why does each user slow the others down? Token generation is limited by memory bandwidth, which all users share. Batching lets one pass through the model serve several requests, so combined output rises, but each user's own speed falls.

Does a longer context mean fewer users? Yes. Each conversation's context is kept in memory. For Llama 3.3 70B, a 32K-token conversation takes about 10.7 GB, so long contexts quickly limit how many users fit.

Is a DGX Spark or GX10 better for a team? Usually, yes. They run Linux with vLLM and SGLang natively, let the GPU use most of the memory, and two units can be linked with ConnectX-7. The GR1X is the better fit if your team wants a Windows desktop for one owner or a small group.