3 articles
Open-source AI is consolidating onto hubs you can build on. LiteLLM is the gateway; under it, Ollama and vLLM are both excellent — 7% apart serially, 4.1× apart under load.
A fully self-hosted pipeline: script to voiceover to render, no cloud in the path. Two mascots, worst-to-best, the 3090/4090/DGX numbers, and why the right model depends on your character.
A year after virtualizing my edge AI supercomputer, I finally measured what the hypervisor costs: single-GPU work is free — the VM even won — and four-way tensor-parallel serving pays a third.
