Screen memory first
If the weight budget already fails, reduce model size or quantization before spending time downloading a checkpoint that will OOM.
Local AI / Memory Fit
Local AI Model Fit: Memory Budget and Quantization Calculator
Check whether a model is worth trying on your memory budget before you benchmark it. This tool deliberately avoids speed guesses and exposes the weight math, safety reserve, and context/runtime reserve.
Browser-onlyGGUF quantizationApple / Mobile / GPUNo speed guesses
Tool
Method
How should you use the result?
Then run a real runtime test
Memory fit only clears the first gate. Throughput, time-to-first-token, long context, concurrency, and thermals require measurement on the target device.
Leave headroom
Mobile and unified-memory systems need space for the OS, UI, KV cache, and other work. The default platform budgets are intentionally conservative.
Calculation boundary
The effective bits-per-weight values are budgeting approximations, not a claim that every GGUF file has identical size. KV cache, MoE experts, vision/audio modules, and runtime workspace can materially change real memory use. Treat the output as a screening tool for deciding whether to continue testing, not as a replacement for real llama.cpp, MLX, Ollama, or app-runtime measurements.
FAQ
Frequently asked questions
How is this different from a basic VRAM calculator?
It combines estimated model weights with a conservative platform safety reserve and the context/runtime memory you choose to leave free. The goal is to decide whether a real load test is worth trying, not to predict speed.
Why do Apple unified memory, mobile shared memory, and dedicated GPU VRAM use different budgets?
Each platform needs a different amount of headroom for the operating system, graphics stack, application memory, and concurrent work. The profiles are intentionally conservative so “fits on paper” is not presented as “runs reliably”.
Why does the tool not estimate tokens per second?
Generation speed depends on memory bandwidth, compute, runtime, kernels, context, batching, KV cache, and CPU/GPU offload. Parameter count and memory alone are not enough for a defensible speed estimate.
Can I use total parameter count for MoE models?
Not as a full performance model. MoE systems require separate treatment of total parameters, active parameters, expert loading, and runtime behavior. Use this result only as an initial conservative memory screen.
Does XBSTACK upload my hardware inputs?
No. The calculation runs in the browser and requires no account. A share link contains only the memory, parameter, quantization, and reserve values you chose.
