Local LLM fit · before the 40 GB download
Parameter count × bits is not enough. Context eats VRAM.
VRAMwise estimates whether a GGUF quant fits on a discrete GPU: weights from published file sizes, KV cache from architecture fields, plus a visible overhead %. Verdicts are FITS / TIGHT / WON'T FIT — not marketing scores.
How this differs from “params × 2 bytes” widgets
- Uses real GGUF size fields per quant when present in our model table.
- Adds KV cache as
kv_bytes_per_token × context / 1e9GB. - Shows headroom so “barely fits” is labeled TIGHT.
- Separate tools for context growth, quant picking, offload speed sketches, and bandwidth ceilings.
Workflow
- Fit checker — GPU + model + quant + context
- Context explorer — raise tokens until it breaks
- Quant picker — Q4/Q5/Q8 vs budget
- Spec tables — raw grids