Dgx spark with Qwen3.8-Flash-Next

MiaAI-Lab

@MiaAI-Lab · Sep 4

Step-by-step recipe to squeeze the 125B model onto one DGX Spark instead of two, with config and measured throughput.

MiaAI-Lab

@MiaAI-Lab · Sep 29

Follow-up engine that speeds single-Spark decoding of the 125B model via a custom tensor layout.

blazux

@blazux · Aug 27

Early write-up of how to serve the model on a single DGX Spark, published within a day of the weights landing.

MiaAI-Lab

@MiaAI-Lab · Aug 26

Dual-node vLLM deployment of the NVFP4 build across a pair of DGX Sparks, for higher context and throughput.

myllmbox

@myllmbox · Sep 1

Compact recipe for single-Spark serving.

myllmbox

@myllmbox · Sep 5

Companion to the solo recipe for a pair of Sparks.

tonyd2wild

@tonyd2wild · Aug 26

VLLM patch that keeps the huge n-gram table on disk and reaches about 44 tok/s on one Spark.

0xBakeer

@0xBakeer · Aug 26

Shows how paging the embedding table to NVMe lets the 180B-class checkpoint run on a single DGX Spark.

vcruz305

@vcruz305 · Sep 7

Serves turboderp's EXL3 quant of the model on one Spark using exllamav3.

sxuff

@sxuff · Aug 26

Tested llama.cpp GGUF route for a single Spark or Asus GX10 box.

azampatti

@azampatti · Aug 30

Convenience script that boots the model on a single DGX Spark or variant.

dolf3131

@dolf3131 · Aug 27

Another route to fit the NVFP4 checkpoint on one Spark by offloading its embedding table.

pangoleen

@pangoleen · Sep 2

Documents a tuning ladder with measured throughput across context lengths.

jschmied

@jschmied · Aug 26

Early notes on fitting the 126GiB checkpoint onto one DGX Spark.

albond

@albond · Sep 2

Experimental one-box DGX Spark deployment.

andreasknopke

@andreasknopke · Sep 14

A stability-focused recipe for a single DGX Spark with a KDA decode kernel and EAGLE speculation, fixing KV-loop failures others hit.

provsalt

@provsalt · Aug 27

Fits the model by quantising the embedding table too.

hashd1ve

@hashd1ve · Aug 26

Day-one single Spark build.

getrefined

@getrefined · Aug 27

TP2 with expert parallelism and MTP speculation.

deathbyorderfill

@deathbyorderfill · Aug 26

Compresses the embedding table so a single Spark holds the whole thing at long context.

bemlerlabs

@bemlerlabs · Aug 31

Hybrid MoE serving on a 128GB Spark.

Weschera

@Weschera · Aug 27

Llama.cpp route with long context.

Showed that UkisAI's reasoning-efficient fine-tune finishes agent tasks at xhigh effort sooner than base Flash-Next at medium, using a Rust/CUDA engine for GB10.