2026-10-17 · America/Los_Angeles · 社区动态 · #14

Ling 3.1 Flash (560B MoE) loaded across 4 DGX Sparks using llama.cpp RPC

sudoing posted that he's loaded Ling 3.1 Flash across four DGX Sparks. It's a 560B MoE with 25B active parameters per token. He'd enjoyed the earlier Ling models and wanted to test this one, but it wouldn't fit on his two Sparks with the setup he was trying. someone in Germany gave him access to a four-Spark setup. He's using Atomic Chat's day-one Q4 GGUF, distributed across all four machines through llama.cpp RPC. According to his post, there wasn't an official FP4 build for vLLM available at the time, which is why he went with this route. No performance numbers in the screenshot yet. He…

热度 44.2 / 100；排名与评分保留该期记录。