localLLMrig
Menu

James Meadlock's GhettoRak: eight DGX Spark nodes

GhettoRak is eight NVIDIA DGX Spark nodes, 128 GB of unified memory on each node, 1,024 GB across the cluster. He sold a 512 GB Mac Studio M3 Ultra after large models ran too slowly to use, and moved from two Sparks to eight.

James Meadlock (J&M Labs). He wrote that we may publish and reuse his material. . Self-reported. Not a localLLMrig measurement.

Test Gemma 4 31B, Recursive Flint and Nemotron 3.5 Super VL side by side in your browser, free.

GhettoRak: an 8-node DGX Spark cluster. GLM-5.3 on eight DGX Sparks: SGLang vs vLLM. DeepSeek V4.1 Flash 1017 on four DGX Sparks.

The cluster

Each node is an NVIDIA DGX Spark with 128 GB of unified memory. Eight nodes is 1,024 GB. The eight-node cluster runs GLM-5.3 at NVFP4, and he has published SGLang against vLLM for that setup, including a unified-memory bug. He calls that run usable.

This week his favourite model is DeepSeek V4.1 Flash 1017, which he calls DSF, on a four-node Spark split. The cluster is split in half right now: four nodes run Imagination Studio, and four nodes run DSF. A four-node Spark split also runs GLM-F or DeepSeek Flash well, and he says that is much better than the M3 Ultra.

He sold the Mac Studio M3 Ultra (512 GB) after Qwen 397B and GLM 5.1 ran so slowly that they were not usable. The write-ups of that machine are on his site, next to the Spark pages.

Would he buy Sparks again?

Yes. He would definitely buy Sparks again. He calls them super flexible. If money were tight he would use a cheap laptop and cloud models instead.

New model hardware requirements and price updates by email. We email a link to confirm the address. Nothing is sent until you click it. Privacy notice.

What would make a speed figure trustworthy

Quants are all over the place, so a benchmark has to say how far the quant sits from BF16 on real work: the output, and the quality of that output. That is how GLM-5.3 at 2-bit can be compared fairly with Qwen3.8-27B at BF16. Prefill belongs in the same report, and so does prefix caching. He would trust a figure that came from a source he trusts, with the test easy to duplicate and every detail public.

Where he looks before buying

Good numbers are hard to find, especially for Apple. After the M5 prices and the capability he saw, he went all in on NVIDIA and grew the Spark cluster from two nodes to eight. The pages he pointed at are the inference-box write-up, the GLM-5.3 SGLang versus vLLM card, and the four-node DeepSeek V4.1 Flash 1017 write-up.

On our list: NVIDIA DGX Spark 128GB, Mac Studio M3 Ultra 96GB, Can I run DeepSeek V4.1 Flash 1017?, Can I run Qwen3.8-27B?.

The numbers on this page were measured by the owner on that hardware. They are not a localLLMrig benchmark, and they do not change the fit scores in the calculator.

Questions

What does James Meadlock run, and on what hardware?
GhettoRak, an eight-node NVIDIA DGX Spark cluster, 128 GB unified memory per node (1,024 GB across eight nodes). The eight-node cluster runs GLM-5.3 at NVFP4, SGLang compared with vLLM, at a speed he can use. Four nodes currently run DeepSeek V4.1 Flash 1017 (DSF) and four run Imagination Studio. A four-node split also runs GLM-F or DeepSeek Flash well. Self-reported, 26 October 2026.
Why did he sell the Mac Studio M3 Ultra?
The machine was a 512 GB Mac Studio M3 Ultra. Qwen 397B and GLM 5.1 ran so slowly that they were not usable. The eight-node Spark cluster runs GLM-5.3 usably, and a four-node Spark split runs GLM-F or DeepSeek Flash much better than that M3 Ultra did.
How is the eight-node cluster split today?
Four nodes run Imagination Studio (https://github.com/drowzeys/Imagination-Studio-by-keys-Wasserman) and four nodes run DeepSeek V4.1 Flash 1017. The four-node DeepSeek write-up is https://al-engr.com/dsv41-1017-four-way-flare-tensorfold-sglang-vllm-4x-dgx-spark.html. The eight-node GLM-5.3 SGLang versus vLLM write-up is https://al-engr.com/glm53-8x-spark-sglang-vs-vllm.html.
What would make a benchmark trustworthy to him?
How far the quant deviates from BF16 on real work output and work quality, so GLM-5.3 at 2-bit can be compared fairly with Qwen3.8-27B at BF16. Prefill and prefix caching in the same report. A public setup that someone else can duplicate.
Would he buy DGX Sparks again?
Yes. He would definitely buy Sparks again, and he calls them super flexible. If money were tight he would use a cheap laptop and cloud models.
May we publish James Meadlock's write-ups?
Yes. He wrote: "You're welcome to publish/reuse any of my stuff." The cluster page is https://al-engr.com/inference-box.html.
Can I test open models in the browser before buying a DGX Spark?
Yes. You can test open-weight models side by side in your browser at https://localllmrig.com/. DeepSeek V4.1 Flash 1017, the model on four of his nodes, is at https://localllmrig.com/can-i-run/deepseek-v4.1-flash-1017.

Find hardware · All builder interviews · Send us your llama-bench result