Savant / Your computer
A GOOD FIT IS MEASURED
Your computer.
Your real workload.
Local intelligence needs to fit the machine and the work. This page helps you prepare for a proper assessment; it does not qualify a model or inspect your device.
BEYOND THE RAM NUMBER
Loading is a start.
Usefulness is the test.
A memory figure by itself cannot establish compatibility, quality or comfort. Proper qualification needs a specific model, runtime, hardware configuration and task.
Compact models may suit everyday assistance. A demanding production-coding task can call for different intelligence. The goal is to help you choose responsibly rather than turn one memory threshold into a universal promise.
- Available memory: leave room for the operating system and applications you actually use.
- Context: include the documents and conversation length the task requires.
- Concurrent work: measure alongside ordinary apps, not only on an idle machine.
- Responsiveness: decide what waiting feels acceptable for the workflow.
- Task quality: check the answer and action against a defined expectation.
- Recovery: observe what happens after an interruption or failed request.
AVAILABLE ON THIS WEBSITE
Plan a local setup.
Enter what you know. “I’m not sure” is a useful starting point too.
Start with what you know
Make a hardware checklist.
Uses only specifications you enter. No system scan or model qualification.
Unsure is a useful starting point. This planner gives you questions to investigate, without a compatibility verdict or model recommendation. Entries stay on this page and are not saved or submitted.
Memory capacity alone is not enough to establish compatibility, response quality or speed. Runtime support, context length, other workloads and real tests matter. Cloud use is a separate, explicit choice with provider terms, privacy considerations and possible charges.
WHEN A CANDIDATE DOES NOT QUALIFY
Keep the limits visible.
The planned setup should report why a candidate is unsuitable: resource pressure, insufficient task quality, unacceptable response time or unreliable recovery.
It may be appropriate to consider another local candidate, reduce the workload, or choose a cloud option. Cloud use must follow your explicit choice. A failed local check should never silently become a remote request.
Is 16 GB enough?
It can be a starting point for compact local models. The model, runtime, context and task still need qualification. This site does not claim that every 16 GB computer can run a particular model well.
Do I need a particular GPU?
No hardware recommendation is established by this planning guide. Qualification depends on a supported runtime, the candidate model and the complete system configuration.
Should I buy new hardware now?
The site does not provide a purchase recommendation. First establish the tasks, supported configurations and measured limits. No qualified hardware shortlist is published here yet.