How it works
This site is a static estimator. There is no backend, no tracking, and no data leaves your device. It matches your browser-reported hardware against a static catalog of local LLMs and their GGUF quantizations, then computes approximate VRAM usage.
Hardware detection
- System RAM — Browsers report your approximate system memory (rounded into broad buckets like 4, 8, 16 GB).
- CPU threads — Detected automatically to understand your compute capacity.
- GPU — Browsers cannot read dedicated VRAM directly. Instead, we use your graphics renderer name and match it against a curated hardware database to estimate your total VRAM and memory bandwidth.
Browsers cannot read dedicated VRAM directly. On integrated and phone GPUs graphics share system RAM, so the estimator treats the GPU ceiling as a fraction of your RAM.
Memory math
Formulas are adapted from standard inference engines to provide highly accurate estimates:
- Weight size = params × (bytes-per-weight of quant) + a full-precision (F32) token embedding pinned to RAM.
- KV cache = layers × numKVHeads × headDim × ctx (F16, 2 bytes/elem, 256-aligned).
- VRAM = (weights − token embedding + KV) × 1.1 — a ~10% overhead for activations and context buffers.
These are upper-bound estimates. Actual usage depends on the runner (llama.cpp, ollama, LM Studio) and context length. Always leave a safety margin; if a model "fits", it should run.
Quantization guide
Rough bits-per-weight for the quants we list (matches real GGUF file sizes):
| Quant | Bits/weight | Use when |
|---|---|---|
| Q4_K_M | 4.89 | default balance (recommended) |
| Q5_K_M / Q6_K | 5.70 / 6.56 | quality-sensitive work, spare VRAM |
| Q8_0 | 8.50 | near-lossless, archival |
| Q3_K_M / IQ3_* | 4.0 / 3.x | tight memory |
| Q2_K / IQ2_* | 3.16 / 2.x | emergency only |
Catalog freshness
The model catalog is generated automatically by fetching configuration data directly from Hugging Face. This ensures we have the most accurate architecture parameters (layers, heads, hidden size, vocab, context) for the newest models.
Limitations
- GPU detection is heuristic; pick a GPU manually in the hardware card if it looks wrong.
- We do not model MoE expert offload to CPU or speculative decoding tokens/sec.
- Parameter counts are the vendor-quoted "B" size; minor ±10% vs raw tensor count.
- No backend — recommendations are client-side math, not a substitute for testing a real run.