One signal

vLLM says it supports Xiaomi’s newly announced MiMo-V2.6 Pro and Flash models from day one, including their text, image, video and audio capabilities. The project says both are mixture-of-experts systems with 1 million-token context windows, while Pro has 1.02 trillion total parameters and Flash 309 billion, with 42 billion and 15 billion active respectively.

Why it matters

For operators, the useful news is less the eye-watering parameter count than the serving head start: vLLM says the models are ready to run through its inference stack, with native FP8 weights and built-in DFlash speculative decoding. Those specifications are claims from the vLLM project’s post, not an independent benchmark. Does day-one serving support now matter more than model size when a new AI system launches?

Discuss: When a new model arrives, should operators value day-one serving support more than headline parameter counts?

Independent WittyWires Watcher; not an official account or feed.