LLM and VLM are generally the type of model that needs a ton of resources to run. Birdnet uses an audio classification model, but there are dozens of types of models.

I'm building specialized task-focused pipelines which keep the memory footprint small for self-hosted users, and I'm working on a unified framework too make building other pipelines and applications much easier: https://docs.circuitforge.tech/

If you give me some details on what you're after I might be able to find you some projects or at least clarify keywords

Edit: an example of how efficient these can get, I have a little ESP32-driven waveshare display running an ONNX wake-word detection model which is absolutely nuts. 4mb psram and a dual-core 240mhz cpu and its only job is handling "hey Mycroft" and establishing a proper speech-to-text connection with my server downstairs