How one developer runs three AI models on a single VRAM-limited workstation

A developer found that loading Whisper, bge-m3, and Gemma simultaneously caused out-of-memory errors on their workstation, which had enough VRAM for only one model at a time. Rather than buying new hardware or splitting workloads across machines, they adopted a sequential loading strategy — loading, using, and unloading each model before moving to the next. The approach works because the tasks do not overlap in time, and the added load latency of a few seconds per model is negligible against longer processing times like a 90-second transcription. Model selection involved trade-offs: bge-m3 was chosen over a smaller embedding model for better Korean-English cross-lingual accuracy, and Gemma edged out LLaVA variants in informal document-image description tests. The setup has run stably for several weeks, and the developer is now working to connect the three models into a unified pipeline for processing calls, emails, and messages.