One binary, three verbs. Chat with a model, put it behind an HTTP API, or measure it — the thing you benchmarked is the thing you shipped.
Paged attention and continuous batching keep the card busy across concurrent requests. A disk-first KV cache carries a session across a restart, so an agent that reuses a long prefix does not pay for it twice.
Vision, audio, speech, image and embeddings go through the same modality pipeline as generation, in the same process. A multimodal app is one deployment, not four.
In-situ quantization of any Hugging Face model, plus GGUF from 2 to 8 bits, GPTQ, AWQ, HQQ, FP8 and bitsandbytes. Ask for a level and it takes a published prebuilt when one exists, or makes one when it doesn't.
CUDA with FlashAttention and multi-GPU tensor parallelism, Metal on Apple Silicon, or plain CPU. Same engine and same commands on each, and a doctor command that reports what it found.
It can call MCP servers over a process, HTTP or a WebSocket. The built-in UI at /ui shows reasoning, code execution and plots inline, and edits branch a conversation with its own state. Pass --no-ui and it is a plain API server.