Latest / Logosofia no Dia a Dia
Serving Your Own Fine-Tuned Model in the Cloud
You've fine-tuned an open-weight model like DeepSeek and now you need to serve it behind an API. This episode walks through the two main patterns — dedicated GPU hosting with weights loaded on an always-on instance, and serverless GPU inference where you pay for compute consumed — and untangles the three very different things "serverless GPU" actually means. We cover which platforms let you bring your own weights (RunPod, SageMaker, Modal, Baseten, Replicate, Fireworks, Cerebrium), what the LoRA hosting shift means for adapter-based fine-tunes, and why cold starts are a bandwidth problem, not…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.