SuperFlux
Research lab advancing high-performance algorithms for large language models.
We build open-source infrastructure at the edge of what LLM tooling can do. Targeting the intersection of modern CPU architecture, SIMD vectorization, and novel algorithms, everything we ship is designed for extreme scale and built to be fast first.
SuperFlux Tokenizer
High-performance Tokenizer
A from-scratch C++ tokenizer built to saturate NVMe throughput on modern hardware. An AVX2-accelerated kernel, lock-free producer-consumer pipeline, and novel merge algorithm targeting terabyte-scale corpora.
More to Come
Research in progress
We're exploring novel approaches across tokenization, inference acceleration, and training efficiency. Follow our GitHub organization for updates as the research matures.
Fast. Novel. Practical.