Open-Weights LLM Capability & VRAM Simulator

The saved inference calculator multiplies seventy billion parameters by four bits, divides by eight and obtains thirty-five decimal gigabytes raw weights. It then adds fifteen-percent overhead and rounds to one decimal, producing forty point three gigabytes. Bars compare raw and source-rounded sizes at ten pixels per gigabyte. Several saved comments and initial labels still mention thirty-five; the actual function applies overhead. These are footprint assumptions, not a loaded model measurement or verified hardware requirement. The tool does not download weights, run inference, query a provider or measure quality. Device and model names are saved labels rather than current specifications or benchmarks. Its cache heuristic multiplies parameters relative to seventy billion, a point zero zero zero one two gigabyte-per-token coefficient, context length and batch, rounds to whole gigabytes, and enforces at least one. With thirty-two thousand seven hundred sixty-eight context and batch one it gives four. Added to forty point three weights this totals forty-four point three, above the entered twenty-four-gigabyte GPU. Bars compare total and capacity at ten pixels per gigabyte. The source divides assumed bandwidth one thousand eight by total and rounds to twenty-two point eight tokens per second, then applies a point-two offloading multiplier and rounds to four point six. Neither rate is measured throughput or a validated architecture model. The edge preset selects thirty-two billion parameters, eight bits, sixteen thousand three hundred eighty-four context, one hundred twenty-eight capacity and eight hundred bandwidth. Weight overhead yields thirty-six point eight gigabytes, cache clamps to one, and total becomes thirty-seven point eight. Bars compare total with capacity at four pixels per gigabyte. Bandwidth divided by total rounds to twenty-one point two, but it is still an unmeasured heuristic. Presets do not replace the stored capability benchmark scores, so their weighted capability index stays ninety point nine. Quantization-loss percentages are fixed lookup constants. Native edge preset, complete JSON export and return to baseline preserve these source limitations without calling any model backend.