Based on the performance of the current open-source models, I don’t see any gain in large models with severely compacted accuracy to deploy on a 2 or 4-node DGX cluster. For example, gemma-4-31B-it with mtp would be a fair choice for a 2-node cluster. However, we cannot deploy another model on it s…

A spark cluster is not really suitable for a team of that size. At the end of the day you are severely limited by the memory bandwidth. A 10 person team, especially if they are devs, will need a lot more capacity than that. I suspect that would even push the DGX station beyond its limits, dependent …

[image] kenleo_lucas: Sparse attention (required by newer large models) is not supported on GB10/SM12X He maybe summarized but did not fact check s..t this statement is highly doubtful Sparse attention (required by newer large models) is not supported on GB10/SM12X M3 with sparse attenti…

honored to be summurized 😄 I just saw glm 5.2 has pretty good result in yet benchmark mostly. I hope it does good job in real life. Probably? 4 units can do 4 bit quant + 256k ctx..? (not sure though) And in Europe, you might find less than10k€ for 2 to 4 units jump. recently I ordered 3600€ x 2 f…

Your summaries are not good, what value does it add to just blindly copy/paste without adding or checking anything? There are benchmarks for M3 which hit 25-30. And you can run that now. The 10 t/s you are referring to is against a different quant, also nvfp4 but possibly slower due to more bf16 la…

That is a really great point! No matter what we see or hear, we need to test and verify it ourselves because we may have different HW/SW envs, especially regarding the operation history. We should collect as much info as we can to get more reliable results (This is the scaling law). All the info fe…

Exactly, show that the model he is using is not good at all. This is the bs I constant see from gemini 3.1-3.5 flash in Ai summary in Google search. Real information mixed with hallucinate bs. He says in summary m3 run at 10ts on 2 sparks in q4. This is complete gaslight and crapshoot. It will never…

Ah llamacpp over rpc, and sub q4. Sorry my bad, I did not take it seriously when looked up. Toy setup imo. By this metric dwarfstarfor ds4f on a single spark also considered a win. The whole point of m3 is massive 1m context. 65k for 1m model? Fine for tinkering but listing as a working recipe? Let…

Maybe take a moment to relax, it seems that reading for full understanding isn’t quite keeping pace with the enthusiasm for criticism and writing rather hasty conclusions. “The whole point of M3 is massive 1M context. 65k for 1M model? Fine for tinkering but listing as a working recipe? Let’s do F…

My bad, sorry guys! I’d love to believe everybody is telling the truth. My point is that whether a model’s decoding speed is 10 or 20 tps doesn’t make a significant difference, at least not to me. “It doesn’t matter whether a cat is black or white, as long as it catches mice, it’s a good cat.” Mode…

I agree that a certain level of token generation does provide a lot of comfort in practical use. However, it’s worth remembering that a model is only as useful as it is reliable. In most cases, we ask questions about topics we are not fully familiar with, which makes truthfulness especially importan…

On the output quality and problem-solving ability, is it worthwhile to get 4 DGX Sparks for coding and production use?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

karol.spark June 21, 2026, 10:17am 15

This is complete gaslight and crapshoot. It will never fit.

Minimax3 on 2 nodes decode ~10.7 tok/s, 4bits - DGX Spark / GB10 User Forum / DGX Spark / GB10 - NVIDIA Developer Forums

Some individuals could benefit from taking more time to think through decisions before relying heavily on AI tools.

Topic		Replies	Views
Qwen 3.8 Flash Next DGX Spark / GB10 Projects	56	6373	August 26, 2026
Concern about the usefulness of a single Dgx Spark DGX Spark / GB10	125	6775	August 12, 2026
Deliberations on 4-sparks cluster advantages DGX Spark / GB10 deepseek , nemotron	90	4333	July 16, 2026
On the Uniqueness of the DGX Spark for Running SotA Models in Unconventional Environments DGX Spark / GB10	90	1830	August 30, 2026
I have ordered a second unit. Don't know why my friends say I'm stupid DGX Spark / GB10	47	4140	May 25, 2026
A Spark to beat M5 Ultra and a MegaSpark to beat 2x Rubin PRO 6000! DGX Spark / GB10 nemotron	50	5021	June 25, 2026
DGX Spark performance DGX Spark / GB10	49	7605	February 13, 2026
Performance characteristics going from 1->2 sparks DGX Spark / GB10	38	969	August 30, 2026
Thoughts out loud! DGX Spark 2 or 4 or 8 DGX Spark / GB10	10	1280	May 11, 2026
Why 273 GB/s? Less Is More, Until It Isn’t DGX Spark / GB10	67	3526	March 27, 2026

On the output quality and problem-solving ability, is it worthwhile to get 4 DGX Sparks for coding and production use?

Related topics