Open methodology

Estimates you can inspect.

The model list is curated; the specifications are not. A build-time pipeline fetches public Hub metadata and applies one consistent, visible formula.

01 / Pipeline

Metadata in. Static pages out.

Before each production build, a Node script requests repository metadata and config.json from the public Hugging Face Hub API. It reads parameter counts, architecture, tags, license, and available weight metadata. The result is cached as local JSON; the browser never calls Hugging Face.

01Hub metadata
02Compute estimates
03Static HTML

02 / Formula

The arithmetic.

VRAM ≈ parameters (B) × bytes/parameter × 1.15 + 1.5 GB
RAM ≈ base VRAM × 1.2 + 1.5 GB
Disk ≈ parameters (B) × bytes/parameter
FP16 · 2.00 B/paramQ8_0 · 1.00 B/paramQ4_K_M · 0.57 B/param

03 / Limits

Capacity is not performance.

These numbers answer “will it fit?” They do not predict exact tokens per second. Real use varies with context length, batch size, model architecture, operating-system allocation, and inference engine—such as llama.cpp, Ollama, MLX, vLLM, or Transformers. Treat the 1.5 GB KV-cache allowance as a useful default, not a guarantee.

MoE models need special care: fewer parameters may activate per token, but the full expert set generally remains resident in memory. We flag them and calculate against total model parameters.