← MRM/Projects/ollama-tune
Developer Tooling · Local LLM Live Rust

ollama-tune

Your local model setup is probably lying to you. Measure it.

Local LLM installs run on defaults nobody checks. The context window truncates at a few thousand tokens and reads as "the model got dumb". The model quietly spills to CPU at larger contexts. The cache setting that is supposed to help makes it slower. None of that shows up unless you measure it on the machine it runs on — so this measures it, then writes the config. The report is free and downloadable now at ollama.theggnetwork.com.

ollama-tune — profile
$ ollama-tune profile qwen3.5:9b GPU AMD Radeon RX 6700 XT · 12.0 GiB Backend Vulkan (what Ollama actually chose, read from its own log) num_ctx on-GPU gen tok/s 32768 100% 47.9 65536 81% 9.4 the cliff [warn] default context silently truncates — the model advertises 262144 [info] ~50 tok/s is memory-bandwidth-bound; kernel tweaks will not move it ✔ recommended: PARAMETER num_ctx 32768

Representative output — not a live instance.

What it does
●Reads what actually happened — GPU, VRAM, and the backend Ollama really selected, taken from Ollama's own log rather than from what the spec sheet implies.
●Sweeps a real ladder — loads the model at each context size, reads how much of it landed on the GPU, and times generation there. The sweet spot is the largest context that still sits entirely on the GPU, measured, not guessed.
●Says when there is nothing to win — generation on this class of card is memory-bandwidth-bound, so the report says so and stops, instead of selling a tune that physics will not deliver. The wins it does find are the context ceiling, the CPU spill and the cache type.
●Applies the fix — the licensed tier bakes a tuned model with the measured context, persists the right environment settings for your OS, A/B tests the cache types by restarting Ollama and always restores the stock setting afterwards, and serves a local dashboard for the whole thing.
●Licensing that can't lock you out — Ed25519, verified offline, one machine per seat with three activation credits for reinstalls and hardware swaps. The updates window is enforced against the build date compiled into each release, not the system clock, so a paid version never stops working and rolling the clock back buys nothing.
●The bug it shipped with first — the original backend detection confidently reported CPU on a machine that was running on the GPU, because llama.cpp logs a "CPU compute buffer" line either way. It read as a clean, plausible result. That is the failure mode I go looking for before anyone else finds it.
Proof
0
tests, Windows and Linux
$49
per seat, one machine, 12 months of updates
0
activation credits per seat
0
calls to the internet — it talks to your local Ollama and nothing else
Rusttokio · clapWindows · LinuxEd25519 offline licensingCloudflare Worker activationSingle binary, no installer
🔒
Live and in use — the free report downloads and runs today, and the licensed tier activates through a deployed worker. The source stays private.

Running models
on your own hardware?

I measure before I recommend, and I say plainly when the ceiling is the hardware rather than the config. Run the free report on your own box first — it shows you what is wrong before you pay me anything.

See it live →