You write the math.
Aether handles the machine.
The AI-first language that compiles one codebase to CPU, GPU, and quantum automatically. From research notebook to production deployment, without rewriting a single line.
AI & Machine Learning
Native tensors, autodiff, @gpu kernels, and built-in experiment tracking.
Cloud Services
Compile to native binaries. 100x faster than Python. Zero cold-start.
Command Line
Single binary tools with parallel execution and instant startup.
Embedded & IoT
4KB binaries on ARM Cortex-M, RISC-V, ESP32. TinyML on-device.
Mobile Apps
Compile to native iOS and Android from a single Aether codebase.
Web (WASM)
Run Aether in the browser with near-native performance via WebAssembly.
parallel { } blocks
Run independent tasks simultaneously with a single keyword. No asyncio.gather, no Promise.all, no goroutines. Dependencies with after(), racing with parallel.race, and automatic cancellation.
The runtime auto-parallelizes independent operations. You opt out of parallelism, not in.
parallel {
users = fetch_users()
orders = fetch_orders()
stats = fetch_stats()
}
// All done here. Use them.
// With dependencies
parallel {
data = load_csv("train.csv")
config = load_config()
model = after(data, config) { train(data, config) } }
@gpu — GPU without CUDA
Write tensor math in Aether. The compiler generates CUDA PTX, ROCm HIP, or WebGPU shaders automatically. Works on NVIDIA, AMD, and Intel GPUs.
No CUDA C++. No manual memory transfers. No thread block configuration. Just math.
def attention(
q: Tensor<Float32>,
k: Tensor<Float32>,
v: Tensor<Float32>
) -> Tensor<Float32> {
scores = (q ** k.T) / sqrt(k.dim)
weights = softmax(scores)
return weights ** v
}
// Compiles to CUDA, ROCm, or WebGPU
// automatically based on available hardware
Start like Python. Ship like Rust.
No types needed to start. Add them when you're ready. Enable #strict for production. The same code grows from prototype to deployment without a rewrite.
data = load("data.csv")
model = train(data)
print("Accuracy: {model.score()}") // Production: add types + strict mode
#strict
let data: DataFrame = load("data.csv")?
let model: Classifier = train(data)
print("Accuracy: {model.score()}") // Same logic. Same file. Just add types.
Unified Compute Model
Write once. The runtime auto-places operations on CPU, GPU, TPU, or quantum processor. No code changes.
100x faster than Python
Compiles to native code via Cranelift/LLVM. Predictable ARC memory management. Zero cold-start.
Python bridge
Use PyTorch, HuggingFace, pandas from Aether. Zero-copy tensor sharing. Zero switching cost.
Formal verification
Mathematical contracts prove your code is correct at compile time. Built for safety-critical AI.
Deploy everywhere
One codebase compiles to server, iOS, Android, WASM, and IoT. No model conversion. No wrappers.
Experiment tracking
@experiment logs everything automatically. Compare runs, track metrics, snapshot models. Built in.