A review of ASUS’s $4,499.99 ProArt P16 RTX Spark says it can run a 27-billion-parameter AI model locally at a median 50 tokens per second. The catch is a demanding setup: the reviewer says the fastest result took two days of Docker and WSL work, plus a smaller model file and a speculative-decoding helper.
Watch Desk analysis
What happened
In a review published by Yanko Design, Ali Gulzar tested an ASUS laptop built around NVIDIA’s RTX Spark N1X chip, with 64GB of shared memory. The test used Qwen 3.8, a 27-billion-parameter model. An 8-bit version ran at 7 to 8 tokens per second in Unsloth Studio; a 4-bit version reached about 14.5 in TensorFold. With a DFlash2 drafter, the reviewer measured a median of 50 tokens per second for chat and about 64 for coding.
Those are results from different model formats and an inference engine configured with a drafter, not a like-for-like speed gain from a single setting. The reviewer says the fast runs were plugged in and used Performance mode, and measured speed rather than answer quality.
Why it matters
The review gives a practical sense of what high-end local inference can look like on a laptop: a model of this size can run without sending prompts to a cloud service, but getting the headline speed takes more than opening an app. The 64GB of memory is soldered, so buyers cannot upgrade it later.
There are trade-offs beyond the price. Under AI load, the reviewer recorded fans at 5,500 to 6,000 rpm, about 57 decibels near the hinge and a hot area above the keyboard. The review also flags compatibility limits for some Arm-incompatible games and x64 applications, so this is not a universal workstation substitute.
Our read
This is a meaningful local-AI demonstration, not proof that every buyer should spend $4,500 on a laptop. The useful takeaway is that a 27B model can reach a brisk reported speed on-device, provided someone is willing to tune the setup and tolerate the heat and noise. If you are considering one, check whether your intended models and tools run on Windows on Arm, and whether you need that level of local capacity at all.
What to watch
- Whether other reviewers reproduce the 50-token-per-second result with the same model, engine and drafter.
- How well the reported speed holds up on battery power and in ordinary sustained work.
- Whether Windows on Arm compatibility improves for the tools people actually use.
Discussion spark: Would you pay a laptop premium for local AI privacy and capacity, or is cloud access the better trade once setup, heat and noise are counted?
Sources and evidence
- Source update (18 November 2026, 16:20 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.