How it works
One build per NPU generation.
Hexagon lets us decide exactly what runs where, so a model is built and pinned separately for v75, v79, and v81. An unsupported operation fails loudly at compile time, before it ever reaches a device.
QHexRT by RunAnywhere
Every modality, 100% on the Hexagon NPU.
Same agentic kernel pipeline as MetalRT, pinned to Qualcomm silicon. Nothing falls back to the CPU.
NPU prefill
12,540tok/s
LFM 2.5 230M · Hexagon v81
TTFT
36ms
LFM 2.5 230M · Hexagon v81
NPU resident
100%
LLM live · other modalities in development
Validation
Checked against the chip, not a simulator.
Every shipped build is verified against the real NPU output, across many inputs and prompt lengths.
Publications
Every claim comes with numbers.
The other engine
MetalRT
Our engine for the Apple Neural Engine. One portable package, validated against the placement it actually gets on the chip.