Most local LLM installs run on bad defaults and nobody notices: the context window silently truncates at ~4k (reads as "the model got dumb"), the model quietly spills to CPU at larger contexts, or a "smaller" KV cache setting actually makes things slower on your GPU. None of it shows up unless you measure on your own machine. This tool does exactly that — no guessing from spec sheets.
$ ollama-tune profile qwen3.5:9b === ollama-tune report: qwen3.5:9b === Hardware GPU : AMD Radeon RX 6700 XT VRAM : 12.0 GiB Backend : Vulkan # what Ollama actually selected, read from its own log Context sweep (placement + generation speed) num_ctx on-GPU gen tok/s load s 2048 100% 49.8 2.1 8192 100% 49.1 2.3 16384 100% 48.7 2.6 32768 100% 47.9 3.1 65536 81% 9.4 6.8 # the cliff prefill : ~273 tok/s over 3218 prompt tokens Findings [warn] VRAM ceiling near 65536 tokens [warn] Default context silently truncates (model advertises 262144) [CRIT] Reasoning model — needs think control [i ] Generation ~50 tok/s — memory-bandwidth-bound; kernel tweaks won't move this Recommended config FROM qwen3.5:9b PARAMETER num_ctx 32768
Real output, real machine. The free report is the whole diagnosis — the paid tier is the one-click fix.
Windows x64 (.zip) Linux x64 (.tar.gz)
macOS build coming. SHA-256 checksums. Single binary, no installer, no telemetry — it talks to your local Ollama and nothing else.
| Free | Licensed — $49 | |
|---|---|---|
profile — full report + recommended config | ✔ | ✔ |
list | ✔ | ✔ |
apply — bake the tuned model + persist settings | — | ✔ |
profile --deep — empirical KV-cache A/B | — | ✔ |
serve — local web dashboard | — | ✔ |
Licenses are per-seat (one computer), verified offline — no phone-home, no subscription lock-out. Every version released inside your 12-month window works forever; renewing (~$29/yr) is only for newer builds. 3 activation credits per seat cover reinstalls and hardware swaps. Air-gapped boxes activate from any browser on any device.
Buy a seat — email for early-access invoice
Card checkout is coming; early-access buyers get their key by email, usually within hours. 14-day refund, no questions.
apply creates a <model>-tuned variant with the measured optimal
context baked in and persists the right environment settings for your OS.
--deep restarts Ollama with each KV-cache type and measures which one is
actually faster on your card — we've measured q8_0 (the "obvious win") forcing a
CPU spill on RDNA2/Vulkan; assumptions lose to measurement. serve gives you a
local dashboard: click to profile, compare runs, apply the winner. It always restores
stock settings, even if a run errors out.
Generation speed is noisy if something else is using the GPU — profile on an idle
card. The Linux/macOS hardware probes are newer than the Windows one. --deep
restarts your local Ollama (and always restores it). The tool tunes the box it runs
on — localhost only, by design.