LLM Speed Simulator

Feel how tokens-per-second, jitter, and first-token latency change the perceived speed of a streamed response.

VRAM Calculator

Estimate VRAM needed to run a model on your homelab.

6.85 GB

weights 3.95 · kv 2.00 · +15% overhead

Parameters

QuantizationQ4_K_M · 4.85 bpw

Context length8192 tok

Overhead15 %

Activations, CUDA context, fragmentation.

Architecture auto-estimated for 7B parameters: 32 layers · 4096 hidden units · 8-bit KV cache. Switch to Advanced to override.

Fits on

RTX 3060 (12 GB)RTX 4070 (12 GB)RTX 3090 / 4090 (24 GB)2× 4090 (48 GB)A100 80GB (80 GB)

Idle

Press “Generate” to stream the source text at the configured speed.