A model that fits in memory by itself may not fit the product. The laptop also needs room for the operating system, document processing, interface, saved data, and ordinary applications the user keeps open.
Choose a pilot task whose complete resource footprint can be evaluated. Narrowing the source and output can be more useful than selecting a large model before the workflow is defined.
Budget the whole execution path
Inventory model loading, runtime buffers, context, output, document images, extracted text, application state, and concurrent work. Keep disk footprint separate from runtime memory.
Use current model requirements and runtime documentation as planning inputs. Exact artefact, backend, context, and hardware determine the experiment; broad parameter counts do not establish application fit.
Scope a fictional private-document pilot
Suppose a small team wants short drafts from selected internal passages. Start with one document type, one bounded source range, one reviewed output structure, and named laptop configurations representative of the audience.
Test cold load, repeated requests, longer supported input, cancellation, another app opening, and failure recovery. Observe memory, responsiveness, completion, and output quality together. No benchmark or hardware minimum is claimed in this proposed matrix.
If the task does not fit, compare a narrower input, smaller evaluated model, reduced concurrency, or a different workflow. Do not silently reduce evidence until the output no longer answers the promised question.
Keep the support decision evidence-based
Define supported, limited, and unavailable configurations from observations. Explain installation and storage separately so users understand what must be prepared before offline use.
Preserve useful document viewing and saved outputs independently where possible. A failed generation should not make the entire workspace inaccessible or erase the source.
The tradeoff is capability versus reach and support complexity. Supporting several model choices can help varied hardware but multiplies testing and customer explanations. One well-understood pilot configuration may be a stronger starting point.
Measure reviewer correction effort as well as technical completion. A resource-efficient task that produces unusable drafts does not establish product value.
The ChalkLens portfolio account provides a local desktop workflow reference. This private-document pilot is fictional and does not publish ChalkLens memory requirements or performance results.
Measure the effect of ordinary document viewing while generation runs, not just an isolated model call.
Questions about memory-sized pilots
Is model file size the memory requirement?
No. Include runtime, context, processing, interface, and concurrent workload, then verify the actual combination.
Should the pilot use only the developer's laptop?
Use audience-representative hardware and state the sample. A powerful development machine can conceal constraints that matter to customers.
What should a feasibility review contain?
Include task scope, resource inventory, exact configuration, observations, quality review, and limits. Bring the packet to an offline AI feasibility discussion.