LocalLLaMA maps non-LLM models for RTX 3060

The interesting takeaway is how much practical local AI now fits on a 12GB card once you stop thinking only in chatbot terms. This thread is less “what can my GPU run?” and more a reminder that speech, vision, and media models are already very usable on mainstream hardware.

• –Speech is the most mature category here: faster-whisper’s published benchmarks show large Whisper variants can run quickly on midrange GPUs while using far less memory than many people expect.
• –Vision workloads are the sleeper hit for local setups, with CLIP or SigLIP for semantic search, YOLO for object detection, MiDaS for depth estimation, and SAM-style models for segmentation.
• –Creative tooling also fits the budget: commenters call out Flux and Stable Diffusion for image generation, Kokoro for text-to-speech, UVR5 for audio separation, and Applio for voice conversion.
• –The real optimization lesson is VRAM discipline, not raw model count; quantized weights and task-specific models matter more than chasing giant general-purpose systems.
• –For AI developers, the upside is privacy and latency: many of these models deliver useful offline workflows without cloud APIs, especially for transcription, media cleanup, indexing, and perception tasks.