Save the model, engine, settings, and hardware.
How MamiLens calculates and tests local AI.
A result is trustworthy only when people can see the inputs, understand the formula, repeat the test and find what remains unknown.
Do not change the prompt between models.
Wrong answers and crashes stay in the record.
Labels show what MamiLens has and has not checked.
What every number means
The listed Q4 artifact size is the starting point. Q5 is planned at about 1.18 times that size. Q8 is planned at about 1.82 times that size. The exact downloaded file remains the final authority.
Longer conversations need more cache memory. The estimate uses the selected context, model family, input type, cache precision and number of people active at once. Concurrent cache is included before deciding GPU fit. Q8 and Q4 cache factors are planning heuristics; the total never falls below the single-request uncompressed allowance. Exact runtime support still needs checking.
MamiLens keeps 3.5 GB to 8 GB for the operating system, engine and background work. Larger computers receive a larger reserve.
VRAM is counted as verified acceleration only when the exact GPU and supported backend are known. Several GPUs are not treated as one simple memory pool.
When the model fits a verified GPU, the calculator divides published memory bandwidth by calculated model size, then applies a conservative 35 to 65 percent efficiency range. This is a planning estimate, not a measured benchmark.
A model must pass memory, input, task and published-context rules. Entered free disk space is also checked in the hardware calculator; when it is omitted, storage remains explicitly unchecked. MamiLens then ranks useful models near the strongest sensible size for the selected computer.
Where the rules come from
1
Record the computer
- Operating system and version
- Exact CPU
- Exact GPU and dedicated VRAM
- Total RAM
- Free storage
2
Record the software
- Engine and version
- Exact model tag or file
- Quantization
- Context setting
- Driver version when relevant
3
Prepare the test
- Use a real task with private details removed
- Save the exact prompt
- Write the expected result before testing
- Include a question the model should not answer when useful
4
Run the model
- Run one warm-up prompt
- Use the saved prompt without changing it
- Record the time before the first word
- Record speed and highest memory when possible
- Save failures and error messages
5
Judge the result
- Was the main answer correct?
- Did it follow the requested format?
- Did it invent facts?
- Was the waiting time useful?
- Could another person repeat the setup?
What each label means
A formula used official specifications and the hardware details entered. Nobody has measured this exact result yet.
A user sent the exact setup and outcome. MamiLens has not confirmed every detail.
More than one independent report supports the same useful result.
The MamiLens team reviewed the evidence and repeated or directly checked the important parts.