A 560B Ling 3.1 Flash on Borrowed Sparks

Ant Ling released the weights for Ling 3.1 Flash, a mixture-of-experts model with 560B total parameters, 25B active and a 1M-token context, @sudoingX reported on October 16. The lab’s previous Flash, released in August, was 124B. On Ant Ling’s own chart, he notes, Ling 3.1 Flash beats GPT-5.6 Sol on 10 of 11 benchmarks and goes 5-5 with Opus 5.

@sudoingX says he trusts the lab because he built on its 3.0 Flash, which ran at 42.1 tokens per second on a single DGX Spark with MTP and powered three builds he shipped. He adds that Artificial Analysis now rates 3.1 at 41, against 20 for 3.0 Flash.

The model would not fit on his two DGX Sparks. So he messaged @niklaslenz_ai, someone on X he had never talked to, and asked to borrow his four. Niklas, who runs local AI at home on 4x DGX Sparks, an RTX 5090, two RTX 5060 Tis and an M5U, agreed.

With a community Q4 build loaded across all four Sparks via llama.cpp, @sudoingX plans to run every task twice: once on the Q4 build locally and once on the full model through the API, with the same prompts, to measure what Q4 quantization costs on a model this large. He expects GGUF to hurt speed and is “doing it anyway.”