pulsar — SSD-streaming inference engine for giant MoE models (Rust + CUDA). GLM 5.2 743B at 2 tok/s and Hy3 295B at 7 tok/s on tw

Pulsar is a new Rust-based SSD-streaming inference engine designed for massive Mixture-of-Experts (MoE) models. It achieves impressive speeds, running GLM 5.2 743B at 2 tokens/sec and Hy3 295B at 7 tokens/sec on just two consumer GPUs by efficiently managing attention and expert layers across available memory. The engine offers zero-configuration multi-GPU support, automatically optimizing placement based on PCIe bandwidth.

Opening Kapyn…