TensorFold is an open-source tool for running large language models quickly on Apple Silicon and NVIDIA GPUs. It uses parallel decoding: a drafter proposes multiple tokens, and the target model verifies them together before committing the matching sequence. The project also uses Metal kernels to keep verification consistent with serial decoding.

TensorFold provides an OpenAI-compatible API, allowing applications to send prompts through familiar endpoints, including chat completions. Users can install it on macOS or Linux, start a model with the command-line interface, and connect a client to a local server. Models are downloaded on first use, and the site offers setup instructions, API documentation, an agent runbook, and implementation notes.

TensorFold shares performance measurements for selected models and hardware, including Apple Silicon systems and NVIDIA DGX Spark configurations. Its published results cover specific recipes and test conditions, with notes that hardware, patches, and methods vary; current-release benchmarks are still pending. The project also highlights collaboration with Mia’s AI Lab on optimisation recipes, setup, and testing