I somewhat agree with the sentiment that you really have to find a fit for model vs use case. I am not a developer or software engineer but rather just as plain generic user compared to most here in the forum. I find Qwen3.6 37B slow for the tasks I want it to do for Hermes and I have been on 35BA3B NVFP4 mostly. The speed matters for Hermes since I use it to typically run schedule checks on emails, make quick queries and reports from SAP etc. Plus I noticed that hermes really eats up context tokens whether the request is simple or moderate.
I do some small QOL app projects every now and then and in that case I would use GPT for it and maybe the spark for quick basic edits.
I think an easier way to look at it is to treat it as you would an employee. Given the models that are available now I treat the spark/agents as more of a staffer or assistant for basic to medium tasks. For “engineer or management personnel” functions, I would still go for frontier. Hopefully there will be a newer model that comes out, around 70-80B with 3-6B MoE. I realized 5.5k prefill and 100tks is already a lot so I’m willing to drop to 3k prefill and 50-60tks for a better model. When that happens I’ll probably treat it as an engineer/management trainee