
We study problems that shouldn't be solvable. Then we solve them.
Realm is a research laboratory and engineering company. We work from first principles, publish the evidence, and turn difficult systems research into tools people can build with.
Rust at the core. Python without a second runtime.
Rust is now OA's primary implementation and API authority. Python mirrors its lowercase modules and PascalCase values: oa::matrix::mat_mul_nt becomes oa.matrix.mat_mul_nt.
The first Python preview exposes Engine and Matrix through the same native Vulkan runtime. Broader Ml, Audio, Vision, and media bindings follow the Rust contracts as those surfaces are admitted—without reviving the old Fn* hierarchy.
| Metric | Rust | C++ | Python | Result |
|---|---|---|---|---|
| Wall time / step | 7.41 ms | 9.19 ms | 9.796 ms | Rust −19.4% vs C++ |
| Training throughput | 8.63K sample/s | 6.96K sample/s | 6.53K sample/s | Rust +24.0% vs C++ |
| Token throughput | 138.10K token/s | 111.38K token/s | 104.53K token/s | Rust +24.0% vs C++ |
| Final accuracy | 92.4% | 92.7% | 92.6% | All pass |
| Generation + reload | Pass | Pass | Pass | Quality parity |
One first Release run per language of the same 300-step Char Transformer: batch 64, sequence 16, all-position training, evaluation, generation, and fresh checkpoint reload. Measured on Intel Iris Xe TGL GT2 with Mesa 26.2.2 and Vulkan 1.4.354. This is a live first-run snapshot—not a fixed-clock, multi-process performance claim. In this run, Rust used 24.4% less wall time than Python.
1let engine = oa::Engine::new()?;2let one = oa::matrix::ones(&engine, [2, 3])?;3let two = oa::matrix::full(&engine, [2, 3], 2.0)?;4let sum = oa::matrix::add(&one, &two)?;56assert_eq!(sum.read_f32()?, vec![3.0; 6]);
Runtime · Engine
One engine. Explicit execution. Vulkan native.
One owner for device, memory, queues, kernels, scheduling, and profiling. Every module borrows the same runtime and completes through explicit events.
OA 0.8.0 is a public development preview. Hardware qualification and performance evidence remain scoped to each named device, driver, and workload.








