On SIngle Spark i am running : GitHub - blazux/qwen3.8-Flash-DGX: recipe for running Qwen3.8-Flash-Next on a single DGX Spark · GitHub and it looks like a clear winner.
So who is the winner of the Duel for now?
GLM or Deepseek?
I think, currently there are not enough “real world” experiences and it may depend on your typical tasks / use cases. At least in tasks, where spatial capabilities count (like controlling Blender via MCP), DeepSeek v4 Flash Vision seems to be stronger.
Also, if speed or context pool size counts, DS4FV wins. For quality, we need more datapoints…
I am using tonyd2wild’s repo.
i’m running qwen3.8 flash next fp8 using aiden stack. will wait for more stable stack of deepseek v4 flash vision before switching.
i just finsihed setting up glm5.3 flash , gona test
Take this for how many grains of salt, but I just got GLM 5.3 flash running stable all day, used deepseekV4 Flash 0731 for what a month (when ever it first came out) before and they are almost identical in real world tasks. I tried them side by side in openwebUI…and maybe GLM has a bit of an edge because it has vision, but Deepseek is also a little faster and you should have a dedicated vision model anyways with sparks. I don’t use benchmarks because I find them useless for real world stuff like Blender, unreal engine etc… but I guess my point is…either one…they are both in the same class.
Yeah me too - I had a good run on GLM overnight, great auto-research, debug, build, ready-application was waiting in the morning. So the moat is clean as expected: GLM for overnight automated burners, Deepseek for interactive more snappy work during the day.
What would be awesome at this point is a way to switch between models much faster. Like having the results of certain computational steps baked into the images.
there is sleep mode in vllm , we haven’t use it much yet , and can be called via API
i can dm , synthetic data but is client related.
The repeating-garbage lock on an agent run is a known one and it is fixable.
If you are on the LibertAIDAI ModelOpt NVFP4 build, or an abliterated one, that checkpoint emits intermittent corrupted token ids. Invisible in prose, but when a bad token lands inside a tool-call block the parser desyncs and generation spirals into the repeating output you saw. That is why it hits a tool-heavy enterprise run and not light single-Spark chat. tonyd2wild traced it to vLLM #54150 and moved the default to RedHatAI/GLM-5.3-Flash-NVFP4, which is compressed-tensors and does not do it. Drop-in, same launcher.
Second, sampling. Set temperature 0.95 top_p 1.0 server-side, z.ai’s agentic profile. Low temperature makes any repetition loop worse, and if the flag is unset the server default is not those values.
I ran 40 enterprise SWE-bench Pro instances on the RedHat checkpoint with those settings, no repetition lock, 36 of 40 resolved. So it is not GLM being bad at enterprise work, it is one of those two.
In my experience running 4x sparks in 2 spark clusters hot for high complexity business automation and face to face work;
ds4flash 0731 beats everything. My dual spark cluster up for almost 6 weeks flawlessly. 24:7 hard work. Not a single issue.
Ds4flash vision - early days. Have switched my two Ds4flash clusters over but seeing very much reduced speed, but native vision is nice and intelligence seems ever so slightly better.
Tried glm 5.3 for 2 days on one cluster and got loops, crap performance and an agent that spoke to me like a robot. Testing it on non-attended dev flows proved it slightly better than ds4flash but slow as hell. Would I surrender a two spark cluster to it? Hell no.
It actually felt like running qwen 3.5 122b on a single spark did - great at coding but crap at communicating and prone to looping. I haven’t had to babysit my agent since ds4flash 0731 and don’t really want to have to start doing that again!.
what quant was it?
Pi. Ctrl+L.
MiaAI Lab EXL3 weights + fp8 KV recipe.
Don’t get me wrong, if your use case is deep research or coding and you need the most accurate model and don’t care about performance than it probably edges ds4flash.
But for most use cases it’s a lot slower and a worse experience to actually communicate with on a daily basis.
Am still not sure how you all are getting the EXL3 variant to loop. I have been running this for days and I don’t think I have seen it loop once, at least not using Entrpi’s engine (well, the first version of it). I have seen the NVFP4 quant do looping, but not EXL3 2 spark quant. @phyo.arkarlwin you mentioned yours was on a CSV file. If it doesn’t have any private data, do you have a copy of it? I’m just legitimately curious as to what’s going on.
I admittedly have not tried the non-quantized versions of DSV4F or DSV4V (or DS4.1(/V)F) on my 2x sparks yet. The closest I’ve come is 0xSero’s single spark quant which I believe is somewhere in the neighborhood of Q3ish, but doesn’t DSV4F also have the capability to loop? Honestly reasoning wise, I don’t really find it to be a small difference. It seems to be a lot more intelligent at the same reasoning level. For coding, I kind of wonder how responsiveness would be rated if most people left it on “high” reasoning instead of “max”. Max has some insane detail retrieval, but it does reason for so much longer to make it happen. Coding is probably not the use case for it.
Same here, ds4 flash 0731 is so good for my workflows.
It accomplishes whatever I throw at it (using for SWE).
And I’ve had almost zero sessions that annoyed me (you know, the ones where you start cussing at the agent - I know it’s not just a me-thing haha). But that’s a very good sign.
And DeepSeek v4.1 Flash is right around the corner. Should be out tomorrow on the API. Hopefully they improve the DSpark head (my intuition tells me that’s the main cause for worst performances on the vision experimental weights). Maybe we’ll get the best of both worlds tomorrow.
Believe me, I know exactly what you mean. I used to get very angry with Claude on a daily basis. I had my fair share of ranty swearing sessions at various Qwen models.
I don’t tend to get angry at ds4flash at all as it mostly just works - follows my workflows, tells me if it can’t do something, and I don’t see the hallucination that a lot of people say is there with the model. Perhaps that’s the way I use it - well defined SOPs, skills that are pruned and fixed.
I love it.
Yeah, I think that 4.1 is basically v4 flash vision exp but in GA form. So - in 2 weeks public weights as always? :D
I was using API Vision Exp for last few days during cluster shakes up - it became 2-3 times faster than in first week it appeared. I think they are bringing changes that come alone with 4.1. I believe its same arch but even more optimized.
Héhé