I went through this recently, I even bought both devices to try. Ultimately I ended up returning both and buying a 5090. In hindsight I’m regretting not buying a rtx pro 5000, but I’m still happy with the 5090.

Mind you, my main use case is agentic programming, so a lot of the model looping and chatting with itself. I’d give it a mostly defined ticket, have it do an attempt, and then review the work and refine it. Currently I’m working on having a more thorough planning step first so that I don’t have to do as much cleanup, but anyway.

I’d give it a decently sized ticket - create CRUD endpoints for new feature X, implement business logic, add database, add unit tests and e2e tests, then review the work and fix any issues. The M4 Max may take an hour, sometimes an hour and a half. The M5 Max would take 30-45 minutes. The 5090 does it in 7-10 minutes, and this is after I limited it to 400W. The difference is huge.

But the other problem, and the second main reason I returned both laptops - the fans spin up the whole time, and the laptop quickly throttles. Imagine hearing the fan blasting for an hour and a half while it does its thing. The M5 Max throttled less, but no difference in fan noise. This also destroys the portability aspect of it - sure, I can take it out of the house, but the battery will die in a couple hours while it’s running. For the 5090 I threw it in a old junker I had lying around, just with an upgraded PSU to support the power requirement, and it lives in the basement that I connect to from my laptop over internet.

Now, with the 5090 I’m able to get Qwen 3.6 27B, Q6_K_XL running, with vision, q8_0 KV and 150k context. Ideally I’d like to bump this up to Q8 and FP16 KV, but that ain’t happening with the 5090, hence my regret with not getting the rtx pro 5000. I’ve not had any issues, but I also don’t know what I’m missing out on. The MBPs would’ve been able to run at this spec, but now it’s even slower because it’s bigger than Q6, and honestly at those speeds it’s not very useful.

One caveat is that I didn’t really bother trying MTP, and while that helps with token generation, it still doesn’t help with prefill. Another one is that we don’t know if a MoE model suddenly drops that somehow beat the 27B, in which case the MBP would be actually able to run it, and get much faster speeds than a dense model. I know 35B only has 3B active parameters, but it was quite usable on the MBPs, it just lacks the reasoning of the 27B.

I still think the ideal is the future M5 Ultra, but who knows what will happen there. We may be waiting for a Nov reveal with shipping starting next year, prices being way higher than the M3 Ultras with less RAM, and being as impossible to get as the M3 Ultras today. I didn’t want to wait that long, and plus worst case I could sell the 5090 if the M5 Ultra ends up being way more reasonable than I expect.

Sorry that ended being more of a wall-of-text than anything helpful, but I hope that at least the small datapoint of a real-world use case helped.