Technology for emerging-market businesses and institutions.
Explore Deviceterra ↗Describe the device. MamiLens will estimate a safe memory budget, shortlist models and recommend a starting engine. It will not invent a speed result.
Understand the words in this guide before making a decision. Start from the beginning, or choose one of the 20 available lessons.
If you do not know the GPU, choose “not sure.” MamiLens will make the safer CPU-only assumption.
These models pass the current memory and input rules. Start with the first recommendation and benchmark it.
A CPU-first setup is the safest assumption when dedicated GPU support is unknown.
More CPU cores can help, but memory bandwidth and engine settings still matter.Qwen 2.5 14B is a curated dense profile for general, writing, coding workloads.
ollama run qwen2.5:14bA compact Microsoft model aimed at strong reasoning, mathematics and instruction following.
ollama run phi4:14bA text assistant with an optional extended reasoning mode for writing, coding, mathematics and tools. Follow its published system instruction to enable reasoning.
ollama run cogito:14bMistral Nemo 12B is a curated dense profile for general, writing, tools workloads.
ollama run mistral-nemo:12bGemma 4 12B QAT is a curated dense profile for general, writing, reasoning workloads.
ollama run gemma4:12b-it-qatQwen 3.5 9B is a curated dense profile for general, coding, reasoning workloads.
ollama run qwen3.5:9bThe catalogue covers 152 current, workstation, mobile, server and important legacy GPU profiles. An unknown GPU or multi-GPU system is never treated as verified acceleration. Every speed range is a bandwidth-based calculation, not a measured result. Exact model files, context, runtime version, background applications and cooling can still change the result.
Choose your RAM, operating system and task. If you have a GPU, select its exact model. Count Apple unified memory once.
Read the memory estimate, execution path and limits. A model can fit in memory and still respond too slowly for your work.
Follow the engine setup, use a short context and try the same few work samples. Keep the model only if the answers and waiting time are useful.
How to test the recommendation →These permanent pages use the same conservative rules as the calculator. Choose the closest setup, then enter your exact device above.
The calculator reserves model weights, runtime memory, operating system memory, model-specific KV cache and extra multimodal memory.
Reported VRAM is counted only when the exact GPU family and backend support are known. Unknown and multi-GPU systems require validation.
A model that loads can still be too slow. Large CPU-only plans are marked as likely slow instead of being called a strong match.
You enter your hardware. The calculator does not scan your computer. Use your system settings to confirm RAM and the exact GPU.
Yes. Small text models can run on a CPU. The result helps check memory; your own test shows whether the waiting time is acceptable.
Results are calculated estimates. Hardware support, memory use and speed depend on the exact model, engine and settings. Read the evidence and limits.
No. System RAM is shared by the operating system and apps. Dedicated VRAM belongs to a graphics card. Apple unified memory is shared by the CPU and GPU.