I’ve had one spark for about six months and have been experimenting with agentic workflows in engineering (langgraph) with custom RAG/MCP for various coding and engineering tool tasks. I also have a 3090 24GB running Qwen3.6-35B-A3B at a blazing ~150 tok/s and a strix halo 128gb unified laptop for on the road demos (etc). I’m looking for insight from others who have gone from 1 to 2 sparks on whether it was “worth it” and what specifically they accomplished by making the move. It seems like these things keep getting more expensive, so I doubt I’d regret the purchase, but I’m looking to justify it…
I recently took the plunge and have no regrets. It’s allowed me to explore larger models, larger context windows, and in many cases (such as Qwen 3.6 27b FP8) significantly improved performance from “too slow” to “genuinely usable”.
Personally, I think 2 nodes allow you to run more models that feel less like a toy and more like an actual tool that’s reliable and trustworthy (DSV4 Flash, Step 3.7 Flash, Mimo V2.5, Minimax 2.7, etc.). Granted, with REAP or smaller quants, 1 node can run them too, but 2 nodes boost them to an actually respectable speed for agentic workflows (30-50 tok/s).
And the scaling from 1-2 nodes is actually not bad; I got 1.7-1.9x speed up going from 1 to 2, but only about 1.4-1.6 from 2 to 4 due to the added latency from the topology.
If you are looking for more reasons… The price tag for your Spark includes the ConnectX interface, which is $1000+. If you paid for it, you might as well use it : )
Yes, I’m also struggling with this issue. After all, it’s not cheap. I’ve always felt that buying a second Spark instance is only necessary if one Spark instance isn’t enough for my use case. Otherwise, just to run higher-quality models, I don’t think it’s very meaningful, even though the ConnectX interface is just sitting there unused.
Yeah, it always comes down to everyone’s use case.
Smaller models like Qwen 27b and 35b a3b are neat, but they just can’t execute my implementation plans reliably (bugs, skipped instructions, etc.) And I HATE that I have to be extra specific with my wording for smaller models, or it might get stuck in a loop trying to figure out what “easier” means (easier for the player, or easier for the player to die?), or do something completely different than what I asked it to do.
I have this issue much much less with 200B+ class models.
I’m traveling right now so I don’t have it in front of me, but with a solo recipe I think I was seeing ~15-16 tps or so. Using a tp=2 setup with Ray, MTP 3 I consistently get about 27 tps which I find very workable for coding tasks. Splitting the tensor work across two Sparks is beneficial despite the networking overhead. And, it does a good job using pi. I always give specific instructions for production code… Very detailed. By choice.
I’ve also had great luck with MiniMax 2.7 NVFP4 and Qwen 3.5 122b FP8. Tried Nemotron 120b but didn’t find it great for coding.
Worth it. I’m not using sparks for agentic development, but as RAG stack (extraction/summarization + vector DB) and researcher agent.
Before was using Qwen 122B, while it was usable, it hallucinated quite a bit and could miss some topic entirely when summarizing.
Dual sparks unlocked Deepseek 4 Flash, which is much better. Also it’s useful as coding agent, almost can one-shot rather complex things.
Also i was able to fit 8B vision model on second Spark (first one has embedder model) alongside Deepseek, after wiring it in LiteLLM i now have multimodal DeepSeek, which is quite neat. :) Added bonus here is that you can finetune vision prompt (stock Qwen encoder produces “amazing, gloruius, expansive image of a landfill” style of descriptions), and make vision actually useful for work.
I have exactly same use case including a vision model 9n a second spark parallel to cluster vllm instance) but I stopped using rag and fully moved to karpath’s wiki, it’s amazing.
One spark for me was on a brink of regret. Two sparks are indispensable now, I literally do 99% of my Ai on them as well as other stuff.
I was considering jumping to 4 but after long research realized it’s a path of diminishing returns in my case. I might take a plunge on m5 max Mac though and run qwen 3.6 27b in parallel as reviewer/validate, minority report Ai to criticize ds4f life choices)
Hah, almost the same here. :) Considering maybe M6 Studio if that is any good to run something 120B-ish very fast.
Actually played with Karpathy-style wiki idea, but it’s not optimal for my use case - very large corpus os transcripts/MoMs and a lot of technical docs, no point in summarizing those. Thinking about graph DB (which is what Karpathy wiki kinda is, just schemaless), but those things are complex.
Quite proud of auto-memory system though - it ingests Claude Code transcripts and extracts facts (this is where DS4 shines compared to 35/122 models) into memory bullets in vector DB, which are queried, reranked and injected into prompts.
It’s like having infinite context, it just remembers everything i’ve done, files i worked, decisions from meetings etc. - all without RAG/MCP roundrtrip (injection itself is 100 ms, automatic). Highly recommend if you use agent chat to drive daily manager-style workflows.
How do you get 150 tk/sec from that 3090 with just 24GB? I am not an AI software engineer, so I have struggled to get anything beyond LMStudio and Ollam running on my sparks (I have 2 of them). However, I was just able to get the two sparks running vLLM, using the recipe at recipes/qwen3.6-35b-a3b-nvfp4.yaml on the 2node cluster. However, it was just generating ~75 tk/sec. I asked the same question using LMStudio on a single node, and it generated 56 tk/sec. My understanding is that vllm inference should be much faster than llama.cpp that LMStudio uses, but I have not seen that. What is everyone doing to get the much better performance?
So just this last week and a half I have clustered two Sparks together after running just 1 for quite sometime. Unlike many I seem to have found a good mix of settings for Qwen3.6-26B using SGLang that runs well, is it a bit slow yes. Everyone’s always chasing the instant response. You can get that but you trade quality. Or you pay for it. There’s a middle ground. Hermes with Qwen 3.6 26B is absolutely awesome. It runs my entire home lab documenting everything to my local “github like product” and its way better than most of the other models I’ve used with it. Gemma gives really fast speeds but lacks the depth I think.
In terms of 2x sparks, it opens the door for a few use cases:
Running much more mid sized models at once for different tasks
Bigger models (duh) is interesting so far as I play with them, yes they are smarter, the speed is ok it depends like any other LLM. Deepseek seems to work well for me.
I have not tried using Qwen 3.6 26B on a cluster, though I think I read here someone saying by having --tp 2 you get about a 10% boost in speed, its not going to be life changing. I can’t say for sure because I’m trying to test out models that don’t fit in one spark as that’s the big win right?
Is it worth it. That question is so loaded with caveats it’s impossible answer. Use case, business case, needs ect…only you can decide if its worth it. For me having one locally at home and 2x clustered at work to play with, I can say I really only need one. Two is a luxury. I I’m not a professional coder, I’m a jack of all trades sys admin, and one spark with a good models (gotta find the right ones for the job) is more than enough even if it takes a little time. I dont need 100 tokens per second, I don’t care if it has to think for a minute before giving me a good answer.
In my humble and very newbie to AI opinion 2x sparks does unlock more ability. The question that’s more important is what are you trying to do with it and is the cost justified.
By the end of the year you’ll be able to buy RTX Spark Laptops. So while I love the idea of having two at home just to tinker with the bigger models, we are also at a weird hardware stage. We don’t know what’s coming in the next year. It may be best to hold off if you can, and see what comes out, if you are a single user imagine what an M5 Mac Studio Ultra with 256 or (hopefully) 512GB RAM unlocks ($$$$). Competition is good for us in terms of HW options in the semi-near future. Instead of buying a 2nd spark for my home I’m putting that money in a safe place gathering interest with zero risk of loss, and continue to add to it as my ‘fun money’ budget, and I’m waiting to see what happens in 6-9 months. See the landscape and then maybe I get a 2nd spark (hopefully cheaper by then) or something more powerful.
Task dictates the model or models you pick.
Your skill and if it’s hobby vs work related dictates hardware and/or paid subscription options.
Knowing how to properly prompt ai to get the best results is very much an art form and a lot of the secret sauce used by the frontier paid models. Of course they have insanely more resources thrown at it than a spark.
Spark shines in concurrency as you likely know already. If all you want is fast fast fast, you can do it on a spark, with small models and their limitations.
It doesn’t sound to me like you are doing this for work, and are interested in more speed than anything else. I would say likely your biggest win will be able to run more mid sized models at once.
But I’m a newb at this thats just my 2 cents, take at face value.