Hey everyone, been watching the community grind on Spark model deploys lately and figured I’d ask the room directly.
I’m trying to put together a short list of what’s actually running well on a 2-node DGX Spark cluster right now, focused on the newer 2026 model drops. From what I’ve seen people land:
Qwen3.5-397B-A17B and MiniMax M2.7 are solid right now
DeepSeek-V4 Flash and MiMo V2.5 have been hit-or-miss
So the question for anyone deeper in this than me: what’s the best model to run for agentic work (Hermes + OpenClaw, tool calling, long context) on a 2-node cluster right now?
Looking for real-world picks based on stability, throughput, and quality on agent workloads, not just paper benchmarks. Drop your current setup + t/s if you have it, would love to compare notes.
Hey, I think it depends on what your agentic work is. If you’re using it to automate your emails/writing or coding, it changes the recommendations.
For agentic coding, MiniMax M2.7 is my current top pick: very good performances on both pp and tg, and better latency/concurrency than with Qwen3.5 397B (which barely fits, which I feel lead to some prefill bottleneck sometimes).
DS4-Flash is excellent through their API, but the ecosystem needs to catch up a bit to run it locally at decent speeds.
I’m also running Minimax 2.7, and it’s a beast when paired with the Hermes agent. That combo is seriously powerful, especially since Hermes is being actively developed—plus, they’re starting to roll out cool add-ons like Hermes Workspace.
Performance-wise, Minimax 2.7 hits about 42 t/s. That’s more than enough for a daily workflow, and it’s consistently ranking near the top when benchmarked against other models.
As for the Hermes agent itself:
In my opinion, it easily beats out OpenClaw.
Nvidia is also planning to drop support for it soon (based on their roadmap and that livestream from late April).
Do we have room/ or is there any way with minimax M2.7 running to still have room for a lightweight vision/stt model? So that the hermes agent/ openclaw can have vision capabilities? Been trying to figure out the best way.
In theory, there should still be room for example Nemotron-3-Nano-Omni. I haven’t personally tested this combo yet, but it should work without any issues.
I agree. Hmmm I currently and juggling Opus 4.7 and 5.5 cloud and then I also juggle kimi 2.6 and the locally going back and forth between 35b, 27b, and minimax. I definitely want to get rid of though Kimi and then try to run 5.5 and opus at the lower tiers but start to implement more and more my local stack. Cron Jobs I notice is what really eats up my usage so even maybe doing that local would be dope but idk.
Do you find that the local models are good enough for your Cron Jobs. like any standout that are really good for them? Or if they all are pretty much good enough, what would be the bare minimum model?
I am testing it now. They are good its more so though when you got them scheduled out right like if you got a ton hititng at the same time that could cause a issue
Yeah I run Qwen 3.6 right now and basically all it does is agentic scheduled tasks. Works like a charm. Probably saves me around $10/day in tokens I’d otherwise have to pay for. If you are constantly running something on your sparks they will pay for themselves within a few years.
I am still really loving qwen3.5 397b across 2 sparks. Its a more well rounded model than minimax in my opinion. I keep trying minimax and it just gives up on fixing issues, and the writing it produces is subpar. Like a brilliant, but very lazy intern. Qwen3.5-397b is like a smart, and very hardworking intern.
I pick the hard worker with enough brain power over the lazy genius.
I just got my second spark, what’s your recipe for 397b if I may ask? I’ve been using Qwen 3.6 27B but I haven’t been super impressed with it (on a single spark). It keeps looping. I hear a few people on “X” saying that the 27B model is better than the larger qwen models because it’s dense. What’s your experience?
To load you cluster to the brim with the largest model you can fit feels tempting, you feel like you get most of your money. But the truth is you often compomise on usefulness in real production and confuse it with experimentation. I caught myself in it again and again. Also generational changes and few months difference are insane now. Older massive model may well be lagging a brand new smaller one. Certainly the case for qwen. Qwen 3.5 122b in any real benchmarks I throw at it is inferior to 3.6 27b. Qwen said 27b is superior to 3.5 398b too.
But in my opinion while 27b is good it is massively brnchmaxxed. Not only it’s not fast, being 27b dense active, but reasoning in qwens is traditionally wasteful.
For my real use, my own benchmarks testing models on polluted data, trick questions stimulated fake news, corrupted data, complex agentic tasks. My current king is Mistral 4 Small 119b A6. 5B. It’s totally working fine on 1 spark, loads 2 million token cache and still 10gb 4am for other tasks. 30 ts. I am of course will try deepseek V4 flash, I tried ds4 iq2xss on one, very good but very slow, but even if it’s usable in terms of speed on cluster, you have to honestly ask yourself if the utility better for you having bit smarter model for tasks most of them can be handled by much smaller model and constantly fight cache availability, oom risk, room for other processes etc. Or having a based faster smaller model with massive cache, stable and much room to spare is better? It’s a point when tinkering should not be confused with production. Don’t sleep on Nvidia Nemotron Cascade 2. It bears 3 Super 9 out of 10, it’s basically a distilled super by nvidia themselves. 60 ts, 20 gb weights, 15gb per 1m q8 tokens. Easily fits in 1 spark with another similar model in a separate vllm. Now I run mistral on one and cascade on another. Both with massive caches, plenty extra ram for agents, harnesses etc. One spark even runs x11 and agent is using visual browser to test pages for script breakage. All comfy
So you feel Mistral is as unbiased as possible / less prone to hallucination?
I find Qwen 397 to be very good at well explained tasks. I mean the model is not perfect but it’s 30tps with fast prefill, full context, large knowledge base, decent agentic skills. It’s the best all rounder and has taken easily to hermes as a harness. Once the memory layer was created, hermes became a personal deep research in my pocket with qwen 397. Minimax was good, but Qwen, subjectively speaking, shows more intent to n the output.
My use is a medical ambient scribe. Qwen receives the transcription, a system prompt and 4 tools: deep_search, searxng, an mcp for ICD-10 codes, wikipedia. Fat Qwen writes a medical note that sounds natural, with impressively correct assessments. Minimax was much more completely factually correct, but short in tone and uninspired. Qwen sounds like a smart colleague, minimax like a bright, but lazy student.
