SparkGLM Serves GLM 5.3 Flash From Two DGX Sparks
@jmacftw's NVFP4 build decodes 65.7 tok/s on code with a 1.37M-token KV pool and 2,760 tok/s prefill.
The SparkGLM team says GLM 5.3 Flash remains the best local model for real near-frontier work on 256 GB, a judgment unchanged since they started the project a month ago. Their build runs NVIDIA's NVFP4 weights at 65.7, 36.3 and 91.6 tok/s for code, prose and structured output on single-stream decode. Cold prefill hits 2,760 tok/s on a custom fork of Atlas, @AtlasInference's engine.
The stack carries 4 x 512K context, 1.37M tokens of KV, NVME disk prefix caching so old long prompts stay reasonably fast, and strict JSON with speculative decoding. The team runs it daily on production workloads and published the repo for anyone to use. Commenters compared notes with the EXL3 TensorFold recipes; @jmacftw concedes those decode a bit faster while his prefill and concurrency win in his use case.