Fix: Meetily is not using your GPU
You have a discrete GPU. Task Manager says the CPU is at 100% and the GPU is at 2%. Meetily is grinding through a ten-minute call and you are watching a progress bar instead of reading notes. This is one of the most-confirmed open bugs on the project, and part of it was fixed in June. The other part was not.
TL;DR
- Transcription got GPU acceleration in v0.4.0, through Vulkan rather than CUDA.
- Summarization still runs on the CPU. The maintainers said so on the thread.
- The Community Edition and the Pro build do not have the same acceleration story.
- Update first. Build from source only if you actually need CUDA.
What people are saying
The sharpest report is meetily#456, opened 2026-05-06 and still open. The reporter had an RTX 3060 with 12 GB of VRAM, CUDA 13.2, cuDNN 9.2.1, and Ollama already running Gemma at 50 tokens per second on the same box. Meetily still took "35 seconds to process 4 seconds of voice on Whisper small-q5_1." Another commenter on 2026-05-11 with the same card called it "completely unusable in this condition" while their i7-9700K sat under full load.
A maintainer answered the same day the issue was filed: "This is something we are working on. However the pro version of meetily does have GPU support now." The next reply, on 2026-05-07, was blunt: "paywalling gpu support is kinda wild."
Then it moved. On 2026-05-20 a maintainer wrote that PR #469 addressed it, "However the GPU usage is not on CUDA but vulkan for more overall compatibility," and on 2026-06-05 that v0.4.0 shipped it. A user confirmed within hours: "Transcription works! However, Summary still uses CPU." The maintainer reply on 2026-06-06 was "This will be addressed in the upcoming releases." As of today, v0.4.0 is still the newest release on the repo.
Summarization on the CPU is its own long-running thread. meetily#333 has 7 reactions and has been open since 2026-02-04: no GPU activity at all while the model generates, confirmed by owners of an RTX 4070 mobile, an RTX 3070 mobile, and a GTX 1050 Ti. And meetily#126, open since 2025-07-03 with 7 reactions, is a user-written CUDA build recipe that predates all of this.
Why this keeps biting you
Meetily is two models stacked on each other. Whisper or Parakeet turns audio into text, then a local model through Ollama turns text into notes. Both are GPU-hungry, they are accelerated by different code paths, and a fix to one does nothing for the other. That is why "did they fix the GPU thing?" has no single answer.
The Vulkan choice is defensible and it is also why your CUDA setup looks ignored. Vulkan runs on AMD, Intel, and NVIDIA from one build, which is the right call for a project shipping installers to strangers. It is not the fastest path on an NVIDIA card, and it does not use the CUDA toolchain you already installed. Meanwhile the shipped binary has to work on machines with no GPU at all, so the safe default is CPU, and the safe default is what you get.
There is also a real cost to building CUDA yourself. A commenter on 2026-06-04 got it working on an RTX 5080 and described running the build "20 times just to unpick all the small ways it was failing," tracing it to an outdated whisper-rs-sys crate that targets sm_52, an architecture CUDA 13 no longer supports.
Fix it in Meetily first
- Update to v0.4.0 or later before touching anything else. Vulkan acceleration for Whisper landed there. If you are on 0.3.0 or earlier, your transcription is CPU-only by design and no setting changes that.
- Update your Vulkan drivers, not your CUDA install. On NVIDIA that means a current Game Ready or Studio driver; on AMD, a current Adrenalin.
vulkaninfo --summaryshould list your discrete card. - Separate the two problems before you debug either. Watch GPU usage while audio is transcribing, then again while the summary generates. If the first spikes and the second does not, you are looking at the open half of the bug and no local change will fix it.
- Move summarization to your own Ollama instance. Ollama does its own GPU acceleration and Meetily can point at it. Confirm Ollama is actually using the GPU first with
ollama ps, which reports how much of the model landed in VRAM. This sidesteps the CPU-only summary path entirely. - Drop to a smaller Whisper model if you are still CPU-bound. Going from
small-q5_1tobaseortinycosts accuracy on technical vocabulary and buys back multiples of speed. On the transcript quality you lose, meetily#474 asks for domain vocabulary hints and is still open, so a smaller model plus a glossary is not yet an option. - Build with CUDA only if Vulkan is genuinely not enough. The recipe in meetily#126 is to edit line 55 of
build_whisper.cmdto add-DGGML_CUDA=1, and the author is explicit that a clean rebuild in a fresh directory is required because an existing CPU build will not recompile. Read the RTX 5080 comment on #456 before you start. - Report your numbers on #456 if none of this helps. Card, driver version, model, and seconds-per-minute-of-audio. That thread got moving because someone posted a measurement instead of a complaint.
Two of those seven steps are "compile it yourself." That is what an open-source local notetaker is, and it is a fair trade if you wanted the fork. It is a strange afternoon if you wanted meeting notes.
Stay or switch
| Job | Stay with Meetily | Switch to LUCI |
|---|---|---|
| You want to swap Whisper, Parakeet, or the summary model yourself | Yes, that control is the product | Partly, model choice is there but the build is not yours |
| You need MIT-licensed source you can audit and fork | Yes | No, LUCI is not open source |
| You want transcription and summaries to use the hardware already in the machine, without build flags | No | Yes |
| You need Linux | Yes, from source | No, Mac and Windows only |
| You want meetings and screen work in one memory your agents can query | No | Yes |
What to do next
I do not think the Vulkan decision was wrong. It was the pragmatic call for a small team shipping to every GPU vendor at once, and the transcription half genuinely got faster. The part that stings is the gap between the two halves, because a user does not experience "transcription is accelerated." They experience a ten-minute call taking a long time.
If you would rather not audit which half of your notetaker is on the GPU this week, LUCI has a different job description. It is not a meeting tool. It keeps your computer's history, the whole day: meetings, email, code, everything in between, with on-device understanding running on the GPU or NPU already in your Mac or Windows machine and Microsoft Foundry Local handling model execution on Windows. Nothing to compile, no flags to pass. Keep the honest limit in view: no build flag makes a two-hour call transcribe faster than the silicon allows. A slow machine is slow for every tool on this page, ours included.
Read LUCI vs Meetily for the full comparison, or the best Meetily alternatives if you are weighing forks.
