ASUS ProArt GR1X for Local LLMs: Why Running a 70B Model Is Only Half the Story

Fitting a 70B model on the ASUS ProArt GR1X is the easy part. Running it well is harder. A 4-bit Llama 3.3 70B file is about 42.5 GB, which fits in either GR1X memory size. But the N1X chip has about 300 GB/s of memory bandwidth, and a dense 70B model must read nearly all its weights for every token. That puts the ceiling at about 7 tokens per second, and published results on similar hardware suggest closer to 4–5. Context length, quantisation and the Windows GPU memory split then decide how useful it is. For most people, a mixture-of-experts model in the same memory will feel several times faster.

Will a 70B model fit, and how much context is left?

Source: file sizes from Hugging Face (unsloth GGUF and Meta checkpoints); KV cache of 0.33 MB per token in BF16 from Llama 3.3 70B's configuration; GPU memory shares as measured on RTX Spark laptops by XDA (64 GB Yoga Pro 9n, 26 Oct 2026) and PCWorld (128 GB ProArt P16, 29 Oct 2026). ASUS hasn't documented the GR1X's GPU memory setting. Context figures assume about 2 GB of overhead and are our calculations; ceilings are bandwidth divided by weight size.

Halving the cache precision (8-bit KV) roughly doubles the context figures, with a small loss in accuracy, according to Ollama's and vLLM's documentation. Our memory sizing guide works through the same sums for 8B, 32B and larger models.

How fast will it actually run?

No GR1X benchmark had been published at the time of writing, so the best guide is the NVIDIA DGX Spark, a GB10 system with 273 GB/s of bandwidth:

• Llama 3.1 70B q4_K_M: 4.4 tokens per second generated, 1,911 tokens per second of prompt processing (Ollama v0.12.6, 23 Oct 2025).
• Llama 3.1 70B FP8: 2.7 tokens per second generated (LMSYS, SGLang, 13 Oct 2025).

Scaling by bandwidth alone, the GR1X's roughly 300 GB/s would put those at about 4.8 and 3.0 tokens per second. That is our estimate, not a measurement, and Windows software may be less optimised than the DGX Spark's Linux stack. For comparison, one RTX PRO 6000 ran Llama 3.3 70B at 4-bit at 32 tokens per second in hardware-corner's tests, which our GR1X vs RTX PRO 6000 post covers.

What does 4–5 tokens per second feel like?

A 500-token answer, roughly a page of text, takes about 100–115 seconds to finish at those speeds. Prompt processing adds a wait before the first word: at around 1,900 tokens per second, an 8,000-token document takes about 4 seconds to read in and a 32,000-token one about 17 seconds, and longer prompts usually process more slowly per token. That suits drafting, overnight batch work and careful review of a few long documents. It is frustrating for chat, coding assistants and agents that make dozens of model calls per task, where waiting minutes per step adds up.

A 70B dense model on a 300 GB/s machine is a capacity demonstration. For daily work, the model's architecture matters more than its parameter count.

Which models feel faster on the same hardware?

Mixture-of-experts (MoE) models store many parameters but read only a fraction per token, which suits bandwidth-limited machines. Techtesters measured these on an MSI Prestige N16 Flip AI+ with 128 GB, the same N1X chip at laptop power (16 Oct 2026):

• gpt-oss-120b (5.1B active parameters): 63.3 tokens per second, sustained at 63.8–65.1 over 10 minutes.
• Qwen3.6-35B-A3B: 78.1 tokens per second.
• Qwen3.5-122B-A10B: 30.6 tokens per second.
• Qwen3.8-27B (dense): 15.8 tokens per second.

Techtesters did not publish its tool or quantisation, so compare within the list rather than with other sources. The pattern still holds: gpt-oss-120b is far larger than a 70B model yet more than ten times faster than our 70B estimate, because it reads far fewer bytes per token. NVIDIA's Nemotron 3.5 Bolt 74B-A9B, packed to 24 GB with 2-bit quantisation, is another option built for this kind of hardware. Our 128 GB memory trap post explains why capacity without bandwidth misleads buyers.

How much quality do you lose to quantisation?

Every step down saves memory and adds speed at some cost to accuracy. Q4 formats are the usual compromise for 70B models. Q2 cuts the 4-bit file by about 40% but brings heavy quality loss, and is mainly useful for experiments. FP8 keeps quality closest to the original but needs a 128 GB GR1X with the GPU share raised to about 96 GB, and it is the slowest option. Test your own prompts at two quantisation levels before settling: the right choice depends on whether your tasks are factual lookups, long reasoning or code.

When is a 70B model on a GR1X worth it?

• You need a specific 70B model, for example one fine-tuned for your industry or language, and no MoE alternative suits.
• You run long, unattended jobs such as summarising a document archive overnight, where speed matters less than privacy.
• One person is the user, not a team. Concurrency on top of 4–5 tokens per second leaves little for each person; see the multiple-users post.

If you need 70B-class models at interactive speed, plan for a GPU with more bandwidth, such as an RTX PRO 6000, or a cloud service. If you can choose the model, start with an MoE model and keep the 70B for jobs where it clearly does better. Our 64 GB vs 128 GB guide helps decide which GR1X to buy, and NVIDIA's 120B and 200B claims explained covers the marketing figures.

Buying in Singapore

The ASUS ProArt GR1X is price on request at SourceIT. ASUS had not published a price at the time of writing; SourceIT estimates about S$6,200–7,200 for 64 GB and S$9,000–12,000 for 128 GB, based on comparable RTX Spark devices, and this is not an official ASUS price. If you want the same chip in a laptop now, the ASUS ProArt P16 128 GB is S$10,773.20 (listed 8 Nov 2026). For 70B models at interactive speed, the RTX PRO 6000 Max-Q is S$29,100 (listed 8 Nov 2026). Prices change; 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 ASUS ProArt GR1X run a 70B model? Yes, at 4-bit on either memory size. A Q4_K_M Llama 3.3 70B file is about 42.5 GB. FP8 (70.6 GB) needs the 128 GB model with more memory assigned to the GPU, if the GR1X allows it as RTX Spark laptops do.

How many tokens per second will a 70B model reach on the GR1X? No GR1X test had been published at the time of writing. Bandwidth sets a ceiling of about 7 tokens per second at 4-bit, and a DGX Spark with similar bandwidth measured 4.4, so expect roughly 4–5 tokens per second.

Is the 64 GB GR1X enough for 70B? For 4-bit with a moderate context, yes. Windows exposes up to about 55 GB to the GPU on 64 GB RTX Spark machines, leaving room for roughly 33,000 tokens of BF16 cache after the weights, or about twice that with 8-bit cache.

What should I run instead of a 70B model? A mixture-of-experts model such as gpt-oss-120b or a Qwen MoE model. On an RTX Spark laptop, gpt-oss-120b ran at about 63 tokens per second in Techtesters' test, many times faster than a dense 70B model.

Does a longer context slow the model down? Yes. The cache grows by about 0.33 MB per token for Llama 3.3 70B in BF16, and each new token has to read it. Long prompts also take longer to process before the first word appears.