Title: Thinking about running Ling-3.1-flash on 4× DGX Spark — how practical is this?
retained text
Ant Ling posted [this demo](https://x.com/AntLingAGI/status/2110390124743970952) of Ling-3.1-flash turning rough briefs into UIs and interactive prototypes. The [weights are on HF](https://huggingface.co/inclusionAI/Ling-3.1-flash) now, and I'm thinking about running it across 4 DGX Sparks. The model card lists 560B total parameters with 25B active per token. Their deployment examples use 8× H200 or 4× GB300, so I'm trying to figure out what a realistic Spark setup would look like. I'd be using it for coding and iterative UI work. Generate something, review it, ask for changes, repeat. Single-user latency matters more to me than batch throughput. What quantization would you start with, leaving enough memory for context and runtime overhead? And which serving stack would you use to split it across four Sparks? my bigger question is whether this would feel usable once it starts making tool calls and reading longer contexts. fitting the weights is one thing, waiting forever between edits is another Would appreciate advice from anyone running large MoEs across Sparks, especially measured prefill/decode speeds or things you wish you'd known before setting it up.
document fetched 2026-10-15 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5uy93/title_thinking_about_running_ling31flash_on_4_dgx