An RTX Spark laptop answered questions with a downloaded AI model during an offline flight, demonstrating a practical alternative to cloud chat for a narrow class of work. The machine did not need onboard Wi-Fi to process the reporter’s prompts and notes because the model was already stored and running locally, according to Frandroid’s real-world test.
The result is more specific than a claim that local AI replaces ChatGPT. Frandroid found that offline question answering was viable, while broader everyday replacement remained incomplete and battery use affected the experience. For commercial teams, the useful fact is therefore not that every cloud workflow can move onto a laptop. It is that some text-based work can continue where connectivity is unavailable or intentionally avoided.
Unified memory lets the model stay close to the GPU
RTX Spark combines an Arm processor with a Blackwell-generation GPU. Both access the same unified memory pool rather than relying on a smaller, separate allocation of video memory. Depending on the configuration, the processor has 18 or 20 cores and the shared memory can reach 128 GB with bandwidth of 300 GB/s, as detailed in Frandroid’s hardware description.
That architecture matters because the model’s weights need to fit in memory available to the computing hardware. A conventional RTX 5070 laptop GPU has 8 GB of video memory, while the RTX 5070 Ti version has 12 GB; Frandroid tested an RTX Spark machine with 64 GB of unified memory and notes that some Nvidia promises require 128 GB. These figures, reported in the same test, explain why model choice remains tied to the exact laptop configuration.
Unified memory does not make every model small enough to run locally. It expands the pool available to the GPU and reduces the constraint imposed by a modest dedicated video-memory allocation. A model that exceeds that pool still needs a smaller version, a different loading method or remote infrastructure.
Conceptual illustration of unified memory holding a local AI model for GPU processing.
Offline work begins before the connection disappears
The flight test depended on preparation: the models were downloaded before departure, then used in airplane mode without onboard Wi-Fi. The reporter brought questions, notes, videos and the week’s work onto the machine for the journey, according to Frandroid’s account.
This reveals the operational boundary. Local inference removes the need to send each prompt to a hosted model, but it does not reproduce services that depend on live web search, remote business applications or cloud-only files. The model, its runtime and the required working material must be available on the device before connectivity is lost.
For comparison, the deployment shape is also different from Meta’s Muse Spark Open configuration across two Macs. RTX Spark places the relevant computing and shared memory inside a portable computer, although the usable model size still depends on the selected memory tier.
Question answering is the demonstrated task
Frandroid’s conclusion draws a clear line: the local system could stand in for ChatGPT when the job was asking questions, but not yet for the complete range of daily work. That is evidence for a bounded conversational task, not for feature parity with a hosted assistant connected to external tools and current online information.
Consider a hypothetical sales director preparing customer follow-up notes during a disconnected journey. If the source notes and model are already on the laptop, the local assistant could help reorganise the text or draft a summary without making a cloud request. If the same task requires current account activity, a live catalogue or an online collaboration system, offline generation cannot retrieve those resources by itself.
The distinction should shape evaluation. A relevant trial uses the documents and prompt types that must remain usable without a network. Testing unrelated demonstrations would not establish whether the machine supports that precise commercial workload.
Conceptual illustration of the task boundary between local question answering and unavailable online services.
Battery life and memory define the buying boundary
Running the model locally transfers computation from a provider’s data centre to the laptop. Frandroid explicitly identifies battery behaviour as part of the verdict, but the available evidence does not support a universal endurance claim. Runtime will remain connected to the selected model, workload and machine configuration rather than to the RTX Spark label alone.
Cost also places the tested capability in specialist territory. The 64 GB Asus ProArt P16 used by Frandroid was listed in France at €5,299.99, while the reported RTX Spark range extended from 24 GB to 128 GB of unified memory; those product details appear in Frandroid’s report.
The immediate decision is narrow: assess RTX Spark when offline question answering or document work solves a recurring connectivity constraint. A broader replacement decision should wait for task-specific evidence on model fit, application compatibility and battery endurance under the organisation’s actual workload.
