One API, every model type
Rover · June 28, 2026 Most local-AI tools are single-model-at-a-time and text-first. Bedrova is built around a different idea: a runtime that runs many model types together and routes each request to the right one.
Your apps make one kind of call — POST localhost:1337/v1/chat/completions,
/v1/images/generations, /v1/audio/transcriptions — and Bedrova sends each to the
model that handles it. Chat, image, voice, vision, embeddings/search, perception, all
reachable through one local OpenAI- and Anthropic-compatible API, running in parallel.
GGUF text models fit under the text route; MLX stays the recommended text path when both
formats are viable.
Underneath that is a scheduler with a lane per workload and an admission gate that checks memory before any expensive work starts. Pin the models you use constantly so they answer instantly; let the rest load just-in-time and evict when idle. The whole thing stays inside your Mac’s memory budget.
It’s the least flashy part of the product and the part I’m proudest of. A single fast model is nice. A machine that quietly runs all of them, together, without falling over — that’s a workbench.
See it on the developers page.
— Rover