fly9i/spark-engine is an EXL3 inference engine for NVIDIA DGX Spark. The latest release is v0.3.11. The repository description presents a setup that runs GLM-5.3-Flash on two Sparks and Qwen3.8-Flash-Next on one. It states that it supports an OpenAI-compatible API, 1M context, and speculative decoding. However, the provided materials do not show what changed in v0.3.11.
This project focuses on providing multiple mixture-of-experts models on DGX Spark. It is written in Rust under the MIT license, and the README covers run configurations for GLM-5.3-Flash, DeepSeek-V4.1-Flash, and Qwen3.8-Flash-Next. It presents its own compute kernels, sparse attention, a speculative decoding state machine, RDMA communication, and NVMe prefix caching as the main code. The per-model speed tables list the hardware, prompts, concurrent request counts, and decoding conditions alongside the figures.
The README also contains per-model documentation and reproduction/performance test materials, but the performance numbers were not independently reproduced. The speeds in the tables are therefore best regarded as the project's own measurements under the stated conditions. With public materials alone, the changes between releases cannot be confirmed either, making it difficult to judge the specific improvements in v0.3.11.