DeepSeek V4.1 Flash on 2 DGX Sparks, and the 2GB memory unlock

On Friday 18 September, one of the regular recipe authors on NVIDIA's developer forum published a way to run DeepSeek V4.1 Flash on two DGX Sparks instead of four.

I had covered the model the week before. DeepSeek released it on 10 September with open weights, and its model card gives it 552 billion parameters. An owner had it running on four DGX Sparks within a day. I finished that story by saying my own four boxes would stay on GLM 5.3 Flash until the recipes settled. A week later they were still changing.

How it fits on two boxes

The recipe works by squeezing the expert layers down to 3 bits per weight, in the EXL3 format, and leaving everything else at higher precision, with the Engram tables still on the SSD.

The author reports 25 to 40 tokens a second for one stream, depending on what the model is writing, and over 50 with several streams at once, serving up to 6 users. That is a little slower than the earlier two-box recipe it builds on, which makes sense for a heavier quant. For comparison, I measured GLM 5.3 Flash on my own four boxes on 11 September at 61.8 tokens a second on code and 29.0 on prose, with one user and thinking switched off.

The 2GB on every Spark

The clever part is a second trick that applies to any Spark. The author says 2GB of every Spark's memory is reserved for the display (4GB if the BIOS setting has been raised), and on a headless box it sits idle. He found a way to hand that memory to the model instead. His recipe uses 1.75GB of the 2GB and keeps 250MB free, which is what the two-box recipe needed to fit its cache.

A few caveats

• As of 19 September I had not run any of this myself. That was the test I had promised the week before, put back because the recipes were still changing.
• The 3-bit quant is the author's own work, and the only quality check at that point was his own perplexity number.
• The Sparks have to be headless, with the display's reserved memory set to 2GB. He says the recipe leaves very little room, with GPU memory utilisation at 0.92, so nothing else should be running on the boxes.
• The speeds are his own measurements, and he says himself that there is no agreed way of counting them.

Read more

• The recipe for DeepSeek V4.1 Flash on two DGX Sparks, on NVIDIA's developer forumforums.developer.nvidia.com
• An owner's earlier report of it on four DGX Sparks, on the same forumforums.developer.nvidia.com
• DeepSeek's model card for V4.1 Flashhuggingface.co

The numbers in this story

My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.