KayonWindows + NVIDIA100% localv1 · preview

Local models, honestly.

Point Kayon at a Windows machine with an NVIDIA card. It reads the hardware, tells you which models will actually run before you spend a gigabyte downloading them, and keeps every token on your own GPU. No account. No cloud. Nothing to trust but the math.

(01)The problem

Not file_size < VRAM. A real memory model.

You download a 13 GB model at 4 MB/s. Forty minutes later it loads, then answers at two tokens a second because it quietly spilled into shared system memory. The tool swore it would fit.

It had compared a file size to your VRAM. Every local-LLM app makes that guess, and it fails exactly when it costs you most. Kayon runs the real numbers per quant instead: weights, KV cache at your context length, compute buffers, and the headroom Windows needs, all against true dedicated free VRAM. You read the verdict for your GPU before the download starts.

Llama 3.1 8B · RTX 4070 Ti · 12 GB
Q4_K_M · 4.9 GBFITS FULLY
Q6_K · 6.6 GBFITS TIGHT
Q8_0 · 8.5 GBGPU + CPU SPLIT
weightsKV cachebuffers
(02)Three bets

What makes Kayon different.

01

Honest fit

One verdict per quant: fits fully, fits tight, GPU+CPU split, CPU-only, or exceeds this machine. The pick that runs best on your GPU leads the page, so the fastest first move is usually the right one.

02

Adopt from Ollama

Already have a shelf of models in Ollama? Kayon links them in place with NTFS hard links. Zero bytes copied, zero re-downloaded. Delete Kayon's link later and Ollama's copy stays right where it was.

03

Private by construction

No account. No cloud. Telemetry stays off until you switch it on, and then Kayon shows you the literal payload before anything leaves the machine. Every outbound request lands in a log you can read.

The core

Explainable down to the gigabyte.

A verdict is never just a color. Open any quant and Kayon shows exactly how it got there.

Weights from the GGUF header, KV cache at your context length (GQA-aware, reading the model's real key_length instead of a naive division), plus compute buffers and CUDA overhead, all measured against true dedicated free VRAM. No shared-memory spill hiding behind a green light.

Verdict breakdown · Q6_K @ 8K ctx

Weights6.3 GB
KV cache @ 8K1.2 GB
Compute buffer + CUDA1.5 GB
Total need9.0 GB
VRAM available11.0 GB
FITS TIGHT — thin headroom
(03)How it works
01

Probe your hardware

NVML reads your GPU, VRAM, driver, CPU, RAM and disk directly, live at 1 Hz. No supported GPU? Kayon still runs and computes verdicts against system RAM.

02

See what fits

Browse a signed, verified catalog. Every quant shows an honest verdict for your machine, and the computed best pick leads. It is never a hardcoded default.

03

Download, checksummed

Resumable downloads with a disk pre-flight and SHA-256 verification against the pinned checksum. A mismatch gets quarantined and never enters your library.

04

Load & chat

One press loads a model through a managed llama.cpp runtime. You get streaming, reasoning, and tool calling wherever the model supports it, with live tokens per second.

(04)Straight answers
No. Kayon is free and runs on hardware you already own. The models are open weights from Hugging Face, so those cost nothing either.
They never leave the machine. Your prompts hit a local llama.cpp server on 127.0.0.1 and stay there. Kayon touches the network only to fetch catalog updates and the models you ask for, and both show up in a log you can open. Telemetry is off until you switch it on, and then it prints the exact bytes first.
There's no sign-in screen, so there's nothing to create. Kayon runs no server that knows who you are.
Kayon finds them and adopts them in place with NTFS hard links. Nothing re-downloads. Delete Kayon's link later and your Ollama blob sits right where it was.
Neither in v1. Kayon runs on Windows with a discrete NVIDIA GPU, and would rather do that one job honestly than half-support everything.
(05)Requirements
OSWindows 10 / 11
GPUNVIDIA · CC ≥ 5.0
Runtimellama.cpp CUDA
FootprintSmall · Tauri

No macOS or AMD in v1. No fine-tuning, no server mode, no multi-agent orchestration. Kayon does one thing: run local models honestly and privately, and do it well.

Stop downloading models that won't run.

Free, offline by default, and entirely yours. Point Kayon at your GPU and find out what it can actually do.