GLM 5.3 Flash on TP4 DGX Sparks (switchless)

[nick398]

@rob01 I haven’t tested DeepSeek 4.1 Flash at all. My last encounter with DSV4Flash was on 0731 version, which I used for some time in production before Qwen3.8-Flash bet it down on my hard math tests. But then GLM-5.3-Flash-EXL3 did beat Qwen3.8-Flash by quite good margin on same test, so I switched to EXL3 version. And above in post you can see FP8 beats it now. So now FP8 version is my production. My work has mix of coding and mostly explorative/frontier work and DSV4Flash-0731 had personality of “worrying too much and fussing about thinking too long”. Qwen-3.8-Flash has personality of “I do my way or you go highway type” where it didn’t give a damn about guidance much (read: forget after few iterations) and followed its genetics/whoever trained it that way to do its way things even though it neither has any world knowledge in then nor can draw parallels (like we humans don) from other transferable concepts. It was daily battle with Qwen3.8-Flash to make it conform to do basic work :) Then came GLM5.3-Flash which is more of explorator than coder in nature. It deliverates a lot (for most of the time good) trying to read every signal/intention/word meaning when its working on frontier work. Maybe its trained that way, maybe it simply does not want to piss me :). But I found from daily use of it much more “open” to follow my guidance on frontier work I am doing. So if I could summarize my own experience with 3 models: GLM5.3-Flash (20 year old graduate of college who looks to problem at hand multiple angles and open for feedback), Qwen3.8-Flash (13 year teenager with his own world) and DSV4Flash-0731 (9 year old kid which is very eager to do things fast but fusses a lot about even smallest hard stuff you throw its way). That is my experience on my own work. Others might have noticed more/less/none of what I noticed.