Can Your PC Run a Local AI Model?

A memory check, not a speed promise

Methodology: The calculator estimates checkpoint weight memory from total parameter count and the selected bit depth. It adds a 10% packing margin, a model-specific working-memory allowance, and a context reserve. Mixture-of-experts models still need their total checkpoint weights available even when only some parameters are active per token.

The estimate follows the practical distinctions described in Ollama's GPU support documentation and the llama.cpp quantization notes. It does not estimate tokens per second, quality, thermals, driver compatibility, or whether a particular backend supports every quantization.

The weight precision selector uses generic bit-width buckets (Q4-Q8, FP16) for memory estimation. Several vendors publish named formats that do not map exactly to these buckets: Nemotron 3.5 Lightning ships NVFP4, Kimi K3 ships MXFP4 weights with MXFP8 activations, Kimi K2.7 Code ships native INT4, and DeepSeek-V4-Flash ships FP8. A user picking 'Q4' is modelling a bucket, not the vendor's published checkpoint.

What the result cannot tell you

Limitations: A passing estimate is not a compatibility guarantee; runtime, driver, backend, thermals, and context behavior can change the result.

• A pass means the memory estimate fits the selected target after a 10% reserve. It does not guarantee that the model will load.
• Long contexts can consume substantially more memory, and the actual KV cache depends on architecture and runtime.
• GPU, unified-memory, CPU-offload, driver, and backend behavior differ. Check the model card and runtime documentation before downloading.
• The model register is intentionally dated. When a release or specification changes, this page should be updated and re-verified.