Intelligence
you own.

QuixiCore: native kernels for every accelerator in the room.

QuixiCore is one kernel contract with six standalone implementations. CUDA kernels are written like CUDA kernels, Metal kernels like Metal kernels, and so on down the line. The backends share the same operations, quant formats, correctness tests, and benchmark method, but no source code.

It covers the operations inference actually spends its time in: quantized GEMM and GEMV, MoE routing and grouped GEMM, paged and MLA attention, speculative decode, and Mamba/SSD. SlimServe's throughput rests on it.

With: a systems language that doesn't fight you.

With compiles to native code through LLVM, with no interpreter and no GC pauses. Ownership-based memory safety rules out use-after-free, double-free, and data races, and you never write a lifetime annotation.

It uses indentation for scope, type inference, f-strings, first-class regex, let … else, scoped with blocks, and direct C interop. The compiler is self-hosting, with CI on Linux, macOS, and Windows for both x86_64 and arm64.

# grep, but pleasant
var nr = 0
for line in file.read_lines("server.log"):
    nr += 1
    if line =~ /error (\d+)/:
        print(f"{nr}: {line}")

fn handle(req: Request) -> Response:
    let user = db.find_user(req.params.id) else:
        return Response.not_found()
    Response.json(user)

let config = with Config.default() as mut c:
    c.timeout = 30
    c.retries = 3