Backed by Y Combinator
A research-first inference lab
We hand-write the GPU and NPU kernels that make consumer silicon fast — and open-source the SDKs, infrastructure, and console that run models on every platform.
0
tok/s · llm decode · m4 max
0.0ms
time to first token
0
tok/s · npu prefill · hexagon
0
open-source sdks · one c++ core
How it fits together
The silicon is already there.
Every consumer device ships with a supercomputer — Apple GPUs, Qualcomm NPUs — that generic runtimes leave idle.
We write the kernels by hand.
No generic abstraction layers. Hand-designed memory layouts, fused operators, engines built per silicon — and every claim published with numbers.
One C++ core above them.
Inference, model management, routing, telemetry — one runtime shared by every platform we target. All of it open source.
Six SDKs, one behavior.
Thin bindings over the same core, so every platform gets the same models, the same API, and the same speed.
Your app, on every platform.
Ship on-device AI to iOS, Android, macOS, Windows, Linux, web, and embedded — and manage the fleet from the console.
your app
ships everywhere
runanywhere-sdks
open source
runanywhere-core
c++ · open source
MetalRT · QHexRT
research · engines
silicon
the hardware
How it fits together
The silicon is already there.
Every consumer device ships with a supercomputer — Apple GPUs, Qualcomm NPUs — that generic runtimes leave idle.
silicon
the hardware
We write the kernels by hand.
No generic abstraction layers. Hand-designed memory layouts, fused operators, engines built per silicon — and every claim published with numbers.
MetalRT · QHexRT
research · engines
One C++ core above them.
Inference, model management, routing, telemetry — one runtime shared by every platform we target. All of it open source.
runanywhere-core
c++ · open source
Six SDKs, one behavior.
Thin bindings over the same core, so every platform gets the same models, the same API, and the same speed.
runanywhere-sdks
open source
Your app, on every platform.
Ship on-device AI to iOS, Android, macOS, Windows, Linux, web, and embedded — and manage the fleet from the console.
your app
ships everywhere
What we do
Inference research that makes the silicon fast. Open-source SDKs and infrastructure that take it to every platform.
01 · Inference — Research
LiveHand-written inference engines for consumer silicon. MetalRT owns the Apple GPU — every Metal kernel written from scratch. QHexRT runs LLM, VLM, STT, TTS, and embeddings 100% on Qualcomm Hexagon NPUs. Every result published with numbers.
0
tok/s decode · m4 max
0
tok/s prefill · hexagon npu
0ms
speech-to-text
658 tok/s decode · 6.6ms TTFT · M4 Max
02 · Open Source — SDKs · Infra · Console
Open sourceEverything above the kernel is open source: one C++ core — inference, model management, routing — with SDKs for Swift, Kotlin, React Native, Flutter, TypeScript, and C++. The hosted console adds OTA model updates and fleet ops on top.
New Research
On-device intelligence - fast, private, hardware-native. Applied research for hardware-native AI inference.
View all publicationsWriting
QHexRTJune 25, 2026
QHexRT is officially live — the first inference engine built to run LLM, VLM, STT, TTS, and embeddings 100% on Qualcomm Hexagon NPUs. First model: LFM 2.5 230M at 12,540 tok/s prefill and 36ms flat TTFT on v81.
MetalRTMarch 15, 2026
MetalRT adds native speech-to-speech support. 1.68s end-to-end latency, 123 tok/s generation throughput, 1.52x faster than mlx-audio on a single M4 Max.
MetalRTMarch 13, 2026
MetalRT adds VLM support and wins every decode benchmark. 279 tok/s vision decode, 92ms time-to-output, 1.22x faster than mlx-vlm across all resolutions on a single M4 Max.
Research-first · Open source
Custom engines for your silicon, or open-source SDKs for your app — either way, it starts with a conversation or a clone.
RunAnywhere Labs
A research-first inference lab. We hand-write the kernels that make consumer silicon fast — and open-source the SDKs and infrastructure that run them on every platform.
Open Source
Research
