NumKong

Handles 15+ numeric types from 4-bit integers to 128-bit complex numbers, with automatic promotion to wider accumulators to prevent overflow, as evidenced by benchmark tables showing zero error for Int8 and minimal error for Float16.

Optimized kernels for x86, Arm, RISC-V, LoongArch, Power, and WebAssembly leverage advanced ISA extensions like AMX and SME, delivering up to 100x speedups over traditional BLAS in benchmarks.

Avoids hidden allocations and thread pools, leaving memory management and parallelism to the caller, which ensures compatibility with arbitrary allocators and threading models, as described in the design philosophy.

Ships as a 5 MB binary, 5-100x smaller than alternatives like PyTorch or OpenBLAS, reducing install size for multi-language bindings across Python, Rust, JavaScript, and more.

Validated against 118-bit extended-precision baselines with compensated summation and saturation arithmetic, minimizing errors—shown in benchmarks where NumKong achieves lower error rates than NumPy or PyTorch for types like Float32.