Local generation uses the customer's device resources. Evaluate energy and thermal behaviour across a realistic session rather than assuming a fast isolated request represents battery use during ordinary work.
Keep task completion and quality fixed when comparing settings. Reducing output or skipping required processing can lower resource use while changing the product's promise.
Record the experimental conditions
Name hardware, operating system, battery or power state, app build, model artefact, runtime, backend, input, output requirements, screen settings, and concurrent applications. Separate cold load from repeated use.
Use current official tools appropriate to the platform, such as Apple's Power Profiler guidance and thermal-state reference. Understand what each observation measures before converting it into a customer claim.
Observe a fictional preparation session
Suppose a laptop imports a chapter, generates several drafts, cancels one request, edits results, and presents a saved pack. Capture resource observations at defined stages and repeat under the same conditions.
Test a longer supported input, low battery conditions where appropriate, another application opening, and idle time with the model retained or released. Verify the interface remains usable and saved work survives any supported adaptation.
No power, temperature, battery-life, or performance result is claimed here. The plan defines the evidence needed before publishing a comparison or hardware recommendation.
Evaluate adaptations without hiding consequences
Possible policies include reducing concurrency, deferring nonurgent work, changing an evaluated backend, or releasing idle resources. Each requires compatibility and task checks. Do not silently switch to a lower-quality mode or cloud path without the relevant product decision and explanation.
Use a clear user state when resource conditions prevent work. A device under pressure needs a safe pause or recovery path, not unlimited retries that worsen the condition.
The tradeoff is timeliness versus device headroom and energy. A setting that completes faster may use resources differently from a slower one. Compare the entire task and repeatability instead of selecting one attractive metric.
The ChalkLens case study documents local desktop inference and classroom workflow. This session is proposed guidance and does not publish a thermal or battery benchmark for that product.
Record whether the device was plugged in and whether the observation includes model loading, active generation, or idle retained resources.
Questions about local AI energy tests
Can token speed predict battery life?
Not by itself. Model load, runtime backend, screen use, task duration, idle resources, and other work affect the complete session.
Should thermal state be treated as an exact temperature?
Use the platform's documented meaning. Do not infer a measurement the API or tool does not provide.
What should a review compare?
Compare equivalent outputs, energy observations, thermal context, responsiveness, and failures under recorded conditions. Bring the traces to a local AI resource review.