bobclawblaw/ripple-65b-nvfp4-m5-ultra is a deployment repository that runs the NVFP4-quantized Ripple-65B-Preview released by ricursiveAI and the DFlash2 drafter on a single Apple M5 Ultra. There is no new version release; what the repository provides is the runtime environment for that configuration and an OpenAI-compatible server. The server supports tool calls, structured outputs, prompt caching, and multiple streams.
According to the repository description, attention, embeddings, and the output layer are kept in 16-bit floating point, while the multilayer parts are stored in 4 bits. This configuration combines quantization that reduces model memory with speculative decoding using a draft model. The execution engine is TensorFold's native Metal engine. It is written in Python and uses the MIT license. The README covers Mac server setup, model download, and the build and run procedures.
However, the README states that the required engine changes are not yet included in the upstream default branch, so a separate patch is used. The speed and accuracy results are likewise presented as the repository's own measurements based on specific hardware and a specific model revision. Therefore this repository is closer to a guide for reproducing that configuration, and its results are hard to treat as performance evidence for other environments.