GLM-5.3 EXL3 Fits Four DGX Sparks

This is an impressive memory-efficiency result, but its practical frontier depends on runtime maturity and real-world quality validation.

• –Keeps the complete MoE topology instead of pruning or merging experts, preserving fidelity at an unusually low 3.38 bpw.
• –Fits weights plus a full 1M-token KV cache within four 128 GB DGX Sparks.
• –Reported serving reaches up to 69 tokens per second with the specialized TensorFold and DFlash2 stack.
• –Stock ExLlamaV3 cannot load the checkpoint, making deployment tooling the immediate bottleneck.
• –Quality evidence is promising but limited; broader coding, reasoning, and long-context evaluations are still needed.