KOLDOS Local

How hardware estimates work

The arithmetic behind the hardware checker, and what it can't tell you.

The hardware checker estimates memory, not speed. It adds up what a model needs and compares it with the memory your system can give it.

Weights

weights = parameters × bits per weight ÷ 8

For KOLDOS 7B at 4-bit: 7 billion × 4.85 ÷ 8 ≈ 4.24 billion bytes, about 3.95 GB. Bits per weight come from common llama.cpp formats; see the quantization reference.

Context

kv cache = bytes per token × context length

Bytes per token depend on the model's architecture, which KOLDOS has not published yet. Until then the checker uses reference values typical for 7B and 22B models with grouped-query attention and a 16-bit cache: 128 KB per token for the 7B and 224 KB for the 22B.

Overhead and reserves

Assumption	Value	Why
GPU runtime overhead	0.6 GB	Driver context and compute buffers the runtime allocates next to the weights.
VRAM kept free	0.5 GB	Memory already used by the desktop, the browser and other programs on the same GPU.
RAM kept free	3 GB	At least 3 GB or 20% of installed RAM, whichever is larger, for the operating system.
Tight fit threshold	8%	Less than 8% of usable VRAM left over is marked as tight: it fits, with little room for anything else.
Unified memory usable by the GPU	67%	On Apple Silicon the GPU can use roughly two thirds of unified memory by default, about three quarters above 36 GB.

Where the model runs

Result	Condition
GPU inference	The whole model and its context fit in GPU memory with room to spare.
GPU inference, tight	It fits in GPU memory, with little room left for other programs or a longer context.
GPU + RAM offload	Part of the model is offloaded to system RAM. It works, but noticeably slower.
CPU inference	The GPU can't hold this model, so it runs on the CPU from system RAM, which is slow.
Does not fit	Not enough free VRAM and RAM combined for this model at this quantization.

The overall verdict

Verdict	Condition
Excellent	The largest KOLDOS model at 4-bit or better fits in VRAM.
Good	A smaller model at 4-bit or better fits in VRAM, but the largest does not.
Limited	The smallest model runs, but with GPU + RAM offload, with CPU inference, or only at 3-bit.
Not enough memory	The smallest model doesn't fit even at 3-bit.

Each result comes with the reasons behind it: usable VRAM and RAM after reserves, what the model needs, and how much would be offloaded. No speed is estimated; the checker shows Performance estimate unavailable until real benchmarks exist.

What it doesn't tell you

Tokens per second. Speed depends on memory bandwidth, the backend and settings, and needs real benchmarks.
Output quality at a given quantization.
Thermal or power limits, which matter most on laptops.
Memory used by other programs beyond the fixed reserve.

Detection in the browser

The Detect button reads the operating system, CPU thread count and GPU name through standard browser APIs (WebGL and user-agent client hints). Browsers don't expose VRAM, so it comes from the published spec when the GPU is recognized. Laptops with two GPUs usually report the integrated one. Nothing is sent to a server.