What are the three terms?

Weights (GGUF size), KV cache, and a runtime overhead percentage.

Frameworks keep scratch buffers beyond pure weights and KV.

TIGHT means total ≤ VRAM but headroom < 10% of VRAM — easy to break with a longer context.

Do you support Apple Silicon fully?

Unified memory differs; discrete GPU VRAM is the first-class path in the fit checker.

Methodology transparency, not a warranty.

Unified memory differs; discrete GPU VRAM is the first-class path in the fit checker.

Methodology transparency, not a warranty.

Every verdict on this site should be reproducible from plain text.

Prefer published GGUF sizes per quant from our model table. If missing, tools that need a size ask you to enter one rather than inventing silently.

From architecture fields when present: bytes per token times context length.

Default 10% on (weights+kv). Increase for vision stacks or concurrent batches.

Exact tokens/sec, multi-GPU sharding, or vendor marketing “AI TOPS” conversions.

General information, not personalized advice.

Models in band

Apple contrast

Runtime tools

Catalog exits

Methodology paths