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.
658tok/s peak · M3
6.6mstime to first token
12,540SDK downloads / wk
6platforms supported
01 / Color
Palette
A warm cream canvas with a signature orange→red gradient for display; green marks passing inference. The gradient is the brand — solid orange carries controls.
Signature — orange → red
--orange
#FF6900
Primary — buttons, links, gradient start
--red
#FB2C36
Gradient end, hot accent, error
--grad
orange→red
Hero display, logo mark
--orange-ink
#D2470A
Orange text on cream (contrast)
Semantic & surfaces
--green
#1B9E77
Passing / success
--bg
#FBF9F9
Warm cream canvas (light)
--bg (dark)
#141210
Warm near-black canvas
--text
#1C1A17
Warm near-black ink
02 / Type
Typography
A high-contrast serif (≈ Newsreader) carries the editorial display; Inter handles UI & body; JetBrains Mono owns every number, kernel name, and the console.
Newsreaderserif · display
from the metal up.
Inter18px · 400
Hand-written kernels that run models fast on every platform.
JetBrains Mono13px · 500
llama-3.2-3b · 96 tok/s · TTFT 90ms
03 / Components
Building blocks
Platform & model chips, inference states, and the CTAs behind the console.
Buttons & model chips
Llama-3.2-3BQwen2.5-1.5BPhi-3.5
Inference state
RunningPassingOOMIdle
04 / Signature
Run anywhere
One model, hand-tuned kernels, running on any device. Pick a target and run — watch tokens stream with live tok/s, first-token latency, and memory.
# one SDK, every platformimport runanywhere as ra
model = ra.load("llama-3.2-3b")
out = model.generate(prompt, target="ios")
Target device
$ runanywhere run llama-3.2-3b --target ios
—tokens / sec
—TTFT
—peak mem
tok/s across platforms
05 / Guardrails
Do & Don't
Do
Use the orange→red gradient for display; solid orange #FF6900 for controls
Darken orange to #D2470A for orange text/links on cream
Set hero & section display in the serif; body/UI in the sans
Keep the inference console dark in both themes
Don't
Don't set body text in the serif — it's display-only
Don't put orange text on the orange button
Don't invent metrics — use real inference language (tok/s, TTFT, kernels)
Don't use Math.random — keep the benchmark deterministic