TyloQuant MFQ
A neural-network quantization and inference toolkit that measures how sensitive each part of a model is, assigns it just enough precision to fit a memory budget, then serves the packed result with CUDA and Metal kernels.
TyloQuant MFQ is quantization and inference infrastructure aimed at running larger, stronger language models within the memory, storage and latency limits of hardware a developer already owns. It combines neural-network-aware scalar and vector quantization formats with fine-grained mixed-precision calibration, measuring activation and loss sensitivity and assigning precision at the level of individual tensors, experts and projections under an exact size budget, then storing the result in a self-contained container format.
A tuned execution path lets a 476-gigabyte DeepSeek checkpoint run directly from its original Hugging Face weights on a single 512-gigabyte Apple Mac Studio, without first converting it into the project's own container format, by keeping the backbone and mixture-of-experts layers resident in unified memory and offloading only a smaller auxiliary table to disk. The formats and runtime are co-designed so that fidelity gained per bit carries through to CUDA and Metal execution speed.
Prebuilt packages are published through the project's releases for Apple Silicon.