SYNAPSE in appatalks/synapse is a Python app that brings local chat, model control, hardware monitoring, and an OpenAI-compatible API together in one place. The repository description sets out the goal of running open-weight mixture-of-experts models (MoE) faster on local hardware. The latest release version is not stated in FACTS. The repository displays an MIT license.
The optional experiment engine is a configuration that splits and runs MoE beyond GPU capacity. Frequently used experts stay on the GPU, while colder experts run on the CPU. A design goal is that the expert contributions selected by the model's router are not discarded, and that target weights are not clipped or requantized. The number of experts, the selection method, and the projection structure differ per model family, which makes a single batching rule hard to apply. The repository explains that it uses a technical description vector extracted from the tensor for batching decisions. This work integrates the upstream llama expert cache and does not claim to be a new idea.
The experimental acceleration is provided in a Linux/NVIDIA research backend, but quality validation is not finished. The throughput basis is limited to three samples on the author's side, and strict-criterion log probability comparisons and answer checks have not yet passed. In one model family, native equivalence also breaks down. Therefore, on current evidence alone, it is difficult to consider the speed goal and output quality as achieved together, and it should be examined separately from the standard execution path.