antirez/ds4 is a C-based local inference engine created by Redis author antirez. Under the name DwarfStar, it supports DeepSeek V4 Flash and 4.1 Flash, DeepSeek V4 PRO, GLM 5.2 and 5.3 and Flash, and Qwen3.8 Flash Next. It covers Metal, CUDA, and ROCm, and was released under the MIT license. It is a project whose scope is limited to specific models.
Rather than broadly running arbitrary model files the way a typical GGUF runner does, this engine targets the handful of models listed in its documentation. It uses SSD streaming when memory is insufficient, and takes Macs with 96GB or more of memory and Strix Halo as its primary targets. It also documents how to set up a multi-user server with multiple GPU cards. It presents configurations that do tensor parallelism on two 128GB Macs connected by RDMA, and configurations that pool the memory of several systems via pipeline parallelism. The README reports an aggregate generation speed of about 126 tokens/second at 8-layer L40S with 16 sessions as its baseline. Support for three backends widens the choice of execution device, but the choice of models stays limited to the scope written in the documentation.
That number is a baseline obtained in its own configuration, not a value re-measured on every commit. No independent verification results are presented, so it does not speak to performance in other environments. Its coverage of several hardware configurations is therefore concrete, but it is more accurate to see this as an engine with a fixed set of supported models and equipment than as a general-purpose runner with a demonstrated performance lead.