The Biggest Mistake When Buying ASUS ProArt GR1X: Choosing Memory Based on Marketing

The mistake is picking 64 GB or 128 GB because a spec sheet says the ASUS ProArt GR1X "runs 120B-parameter models", without checking what your own models need. The real requirement is the model's weights at the precision you'll use, plus the memory for the context (the KV cache), plus headroom. And on Windows the GPU doesn't get all of the system's memory. In practice, 8B and most 32B models at 4-bit fit comfortably on the 64 GB model. A 70B model at 4-bit is borderline on 64 GB and comfortable on 128 GB. gpt-oss-120b needs the 128 GB model and a larger-than-default GPU memory share.

What does each model size actually need?

Source: weight sizes from Hugging Face model files (accessed 6 Nov 2026); context memory calculated from each model's published configuration with a BF16 KV cache, one conversation, before engine overhead (our arithmetic). GPU-visible memory: XDA (26 Oct 2026) found at most about 55 GB exposed to the GPU on a 64 GB RTX Spark laptop; StorageReview and PCWorld found a default split of about half on 128 GB machines. These are laptop findings; the GR1X's own memory settings weren't documented at the time of writing.

Why isn't 64 GB or 128 GB the number that matters?

Because Windows doesn't let the GPU use all of it. On RTX Spark machines, Windows divides memory into a CPU share and a GPU share set in advance. Linux GB10 systems such as the DGX Spark work differently. XDA found that on a 64 GB Lenovo Yoga Pro 9n, the most Windows reported for the GPU was about 55.4 GB, which is why gpt-oss-120b (about 65 GB) doesn't fit on a 64 GB unit. On 128 GB machines the default split is about 64 GB each. PCWorld reported that ASUS's MyASUS app on the ProArt P16 laptop lets you give the GPU up to about 96 GB.

Whether the GR1X has the same setting wasn't published at the time of writing, so check before you rely on it. Even at 96 GB, Windows, your apps and the inference engine all need memory too. Size against what the GPU can actually use, not the number on the box.

How much does context length add?

A lot, and it's the part marketing usually skips. Every token of conversation history or document you feed in is stored in the KV cache, and that grows with context length:

• Llama 3.3 70B needs about 0.33 MB per token. At 32K tokens that's 10.7 GB; at 128K it's about 43 GB, which takes a 4-bit 70B model from 53 GB to about 85 GB. Using an 8-bit KV cache roughly halves that.
• Qwen3-32B needs about 0.26 MB per token: 8.6 GB at 32K and 34 GB at 128K.
• gpt-oss-120b is unusually light, because only half its layers keep a full-length cache: about 1.2 GB at 32K and 4.8 GB at 128K.

That's why ASUS's claim that the GR1X "runs 120B-parameter LLMs locally, up to 1M tokens of context" needs reading carefully. ASUS gives no precision or KV-cache settings. By our arithmetic, gpt-oss-120b with a full 1M-token BF16 cache would need about 104 GB, more than the roughly 96 GB GPU share seen on 128 GB laptops. With an 8-bit cache it would need about 85 GB. It's possible under the right settings, on the 128 GB model, with a raised GPU share. Our post on what NVIDIA's 120B and 200B claims really mean goes through the assumptions behind those headline numbers.

Buy memory for the largest model you'll run, at the precision you'll run it, with the context you actually use. The headline model size on the box is a ceiling, not a recommendation.

Does more memory make models faster?

No. Memory decides what fits, and memory bandwidth decides how fast it runs. Both GR1X models share the same N1X chip and, by reviewers' calculations, about 300 GB/s of bandwidth. A dense model has to read all of its weights for each token, so the theoretical ceilings are roughly:

• about 61 tokens per second for an 8B model at 4-bit (4.9 GB);
• about 15 tokens per second for a 32B model at 4-bit (19.8 GB);
• about 7 tokens per second for a 70B model at 4-bit (42.5 GB).

These are our upper-bound calculations; real speeds come in lower. Mixture-of-experts models read only their active parameters per token, which is why they run much faster than their total size suggests. On an MSI Prestige N16 Flip AI+ laptop with the same chip, Techtesters measured about 63 tokens per second on gpt-oss-120b and 30.6 on Qwen3.5-122B-A10B, but 15.8 on the dense Qwen3.8-27B (16 Oct 2026, quantisation not published). Our post on running a 70B model on the GR1X looks at what dense-model speeds feel like in use.

So which memory should you buy?

• 64 GB is enough if you mainly run 8B–32B models at 4-bit or 8-bit, image generation, coding assistants or agents built on mid-size models. It can't hold gpt-oss-120b, and 70B at 4-bit leaves almost no room for context.
• 128 GB is the choice for 70B-class dense models with useful context, gpt-oss-120b and other 100B-class mixture-of-experts models, or several mid-size models loaded at once for agents. Remember the memory is soldered, so you can't upgrade later.
• Neither holds 200B-plus models such as Qwen3-235B at 4-bit. For those you need a Linux GB10 pair, a larger unified-memory system or multiple GPUs.

For a workload-by-workload comparison, see GR1X 64 GB vs 128 GB. If several people will share the machine, the context memory multiplies per user, which we cover in whether the GR1X can serve multiple users. And the 128 GB memory trap explains why capacity without bandwidth has limits.

Buying in Singapore

The ASUS ProArt GR1X comes in 64 GB / 1 TB, 128 GB / 2 TB and 128 GB / 4 TB, 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$6,200–7,200 for 64 GB and S$9,000–12,000 for 128 GB, including GST. The step from 64 GB to 128 GB is the biggest price jump, so it's worth sizing carefully. For comparison, the ASUS Ascent GX10 64 GB / 1 TB on Linux is S$6,670 and the 128 GB ASUS ProArt P16 RTX Spark laptop is S$10,773.20 (both 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

Can the 64 GB GR1X run a 70B model? At 4-bit, the weights alone are about 42.5 GB, and with a 32K context the total is about 53 GB. That's right at the roughly 55 GB the GPU could use on a 64 GB RTX Spark laptop, so it may load with a short context but leaves almost no headroom. The 128 GB model is the safer choice for 70B.

Can the 128 GB GR1X run gpt-oss-120b? The weights are about 65 GB, which is more than the 64 GB default GPU share seen on 128 GB RTX Spark laptops. It fits once the GPU share is raised; ASUS's ProArt P16 allows up to about 96 GB. Check the GR1X's settings before you rely on that.

How much memory does context use? It depends on the model. Llama 3.3 70B needs about 0.33 MB per token, so 10.7 GB at 32K tokens and about 43 GB at 128K. gpt-oss-120b needs much less, about 4.8 GB at 128K.

Can I add memory to the GR1X later? No. The LPDDR5X memory is soldered. Only the SSDs, in two M.2 slots, can be changed.

Does 128 GB make the GR1X faster? No. Both models use the same chip and memory bandwidth. 128 GB lets you load bigger models and longer context, but a given model runs at about the same speed on either.