exact arithmetic: Nik writes its own working; whenever it writes "expression =", the engine fills in the exact digits (a pencil that cannot miscalculate). Nik still decides what to compute and writes every other word. thinking paths: when Nik reasons (socheko), it can think several times independently and answer with the consensus (self-consistency). Slower, more reliable. ar: one token per step through a hand-written WASM-SIMD int8 kernel, the fastest on CPUs. ssbd (self-speculative block diffusion): Nik drafts a whole block in parallel with its diffusion head and verifies it with its own AR head. It produces the same text in fewer sequential steps. Experimental here; it pays off on GPUs.
A tiny language model for romanized Nepali, built from scratch by Nikit Hamal. I think before I answer, and every word is generated on your device. No server, no API.