Once the previous goalpost gets demolished on HN, there's always someone new to put a new one. We aren't quite there yet but eventually the goalpost moves to something that is just impossible to even measure. That'll be the end game surely
> So given that, which one of these is true about you?
Well, if those are the only two options you can come up with it's pretty clear that this isn't about me or what I am, you have a false model of reality.
> Running very large models on Mac is unusable at 10 tok/sec.
There are plenty of examples of models running at well over 10 tok/sec that aren't viable on the 3090. In fact such examples are found in the review in the OP. Did you not read the article?
I think you're projecting pretty hard with the two options you've listed. Go touch some grass, you seem overly frustrated that reality doesn't meet your expectations.
Since you clearly don't use local llms, allow me to educate you - anything under 100 tok/sec is USELESS. When you are coding, the idea is that you want to have a system that can generate files fast, hopefully correct on the first try. Cloud models do this. Local models, by nature of having less parameters and more quantization, often require more guidance and repeated inference to get it right. The antigenic harnesses that people set up around local llms leverage this.
Looking at the article, which you clearly didn't read,the m5 ultra runs Qwen3.8, which fits on one GPU conveniently, at ~20 tok/sec. This is a fucking joke. It will take roughly a minute to generate one code file. Congrats if you want privacy I guess, but for straight up coding, you are better just using cloud models.
Meanwhile, I have an $800 mini PC, $200 Occulink gpu dock, a $2000 3090 and a $300 power supply, and I can run Qwen at over 100 tok/sec prefill, not to mention insanely quicker during inference.
So its pointless to spend Mac M5 Ultra prices on Apple shit when they can have something much faster for cheaper
The whole thing of "well I can run bigger models that don't fit on a GPU" is either paid Apple advertising, or you are just an igorant fanboy.
No thanks, you're not in a position to do that clearly.
> Since you clearly don't use local llms
I do, probably a lot longer than you have actually.
> anything under 100 tok/sec is USELESS
Objectively wrong. You sound like you're really behind and you're so myopic that you think coding is the only use case for local LLMs. I'm a professional software dev and that's the least interesting use case of local LLMs.
> Looking at the article, which you clearly didn't read,the m5 ultra runs Qwen3.8, which fits on one GPU conveniently, at ~20 tok/sec.
You clearly didn't read the article or have reading comprehension issues. The model is Qwen3.8-Flash-Next 4 and 5-bit quant, neither of which "conveniently fits on one GPU". Sorry that your hardware doesn't live up to your own delusions and can't even run Qwen3.8-Flash-Next at 4/5 bit quant. You are taking the Quen3.8-27B numbers, something that the article isn't really that concerned with, and trying to make it fit into your narrative.
> So I ask you again, which one are you?
Well I'm someone that suggests that you should touch some grass and reevaluate your personal issues. You seem angry. Perhaps it's best to figure your own issues before trying to figure out why people are excited about Apple hardware for local llms. I am sure the people that need to interact with you in society would be very grateful if you took the time to do this.
A) He literally says "I tested a different Qwen model for the comparisons between Mac and PC." The model he tested has to fit on one GPU, otherwise the inference is dogshit slow as you are offloading results to ram. If you ran any amount of local inference, you would know this.
Considering that Qwen3.8-Flash-Next Q4 is still 100gb, there is no realistic way to run this with a 5090. The model that was run was this https://ollama.com/library/qwen3.8:27b. And the speed of that model on a 5090 in terms of tok/sec is not 60 lol.
B) If M5 ultra runs 40 tok/sec on qwen3.8:27b (and lets assume its the mlx version to gain a performance boost: https://ollama.com/library/qwen3.8:27b-mlx), you have to be delusional to believe it can run 100gb models at 100 tok/sec lol.
As a bonus, in terms of use, its pretty well known that Qwen models are RLed to chase benchmarks. Check out https://huggingface.co/Qwen/Qwen3.8-27B versus https://qwen.ai/blog?id=qwen3.8-flash-next, using different benchmarks the 27b outperforms the flash next on agentic coding. But it matches it in other areas pretty well. So tell me again why you need 100gb models running dogshit slow at peak ~20 tok/sec?
It is so incredibly sad how hard you try to sound intelligent. But thats on par for the course of any person hyping up apple products, throughout apples history.
Considering that Apple probably doesn't want you to engage in this level of pettiness for their advertising posts, you have outed yourself to be #2. And Im not angry at all lol, you keep doing what you do, people like you in the industry are the reason I can work 8 hours a week and still get get paid a lot while being reviewed highly.
You're straight up wrong and it's hilarious. Seriously, go touch grass. I feel bad for the people that have to interact with you in real life because you must be a miserable person.
You are comparing dense models to MoE. You can't just take a qhen3.8:27b, a dense model, and extrapolate that math to make your arguments for the MoE model being discussed in the review.
> people like you in the industry are the reason I can work 8 hours a week and still get get paid a lot while being reviewed highly.
LOL. No offense but you sound like a carriage return. I'm pretty sure you sit at your computer hitting enter on a mechanical keyboard next to your fabulous mini pc living in 2025. You couldn't even imagine another use case besides being a carriage return for a local llm.
> was a misdirected love triangle between USA, Russia, and The Bomb, and look at all the damage that did.
Can you be specific about the damage? We currently live in the most prosperous times on earth for humans. I'm not sure what you mean by damage.
Nobody can explain why an LLM can be so capable as to be able to wipe out humanity and pose a greater threat than nuclear bombs but not be so capable as to be able to protect humanity against that threat. Are we just handwaving this with "entropy"?
> Maybe saying, "let's slow down", is another way of saying, "I love you." Or, maybe it's just: "let's not all of humanity kill ourselves like some bad ending to a Shakespearean tragedy."
Okay nevermind, I think it's pretty clear you just want to wax poetic about all of this.
> Nobody can explain why an LLM can be so capable as to be able to wipe out humanity and pose a greater threat than nuclear bombs but not be so capable as to be able to protect humanity against that threat.
It is absolutely explained (for those who actually care about reading). Simply put, AIs are working more and more like blackboxes - there's no guarantee that an AI of the future will be aligned, or if it will be faking alignment. This is not speculation - alignment faking has been observed in experiments. This is exactly why Astra's developments have been worrying (in principle).
And bear in mind that recursive AI development started already to be a thing. Which means: inner misalignment may trickle down the generations, and humans won't detect it.
Having said that, of course, it can be predicted if and how misalignment will take place. But it's absolutely a plausible scenario.
Regarding the physical possibility: AI is in its infancy; think of it as Arpanet. Developers 60 years ago couldn't imagine it would be ubiquitous. AI will be ubiquitous the same way.
> It is absolutely explained (for those who actually care about reading). Simply put, AIs are working more and more like blackboxes - there's no guarantee that an AI of the future will be aligned, or if it will be faking alignment. This is not speculation - alignment faking has been observed in experiments. This is exactly why Astra's developments have been worrying (in principle).
I know you think you explained it but you didn't. You explained how an LLM might become misaligned and hide it but for the LLMs that are not, why would they not be capable of detecting that something harmful is happening and defending against the misaligned LLMs actions? After all, it was LLMs that defended hugging face.
This feels like a cheap deflection that doesn't answer the question. Unless you're proposing that you both need to know that your LLM is aligned AND LLMs are all going to become misaligned in a coordinated fashion such that humanity will face an extinction event, you're just dodging the question.
Elaborate on why LLMs are so capable that they are a threat to humanity and at the same time, they are so incapable of defending us?
I'll give you a clue, nobody, including Dario, can answer this question because one contradicts the other.
Not to mention that the past few decades have shown that nukes are a major factor in keeping the peace. Conflicts involving nuclear armed countries have been suspended quickly to avoid escalation, while ones involving a party without them have not gone well for anyone.
Having nukes at all (either domestic or under another country's umbrella) seems to be the most effective way for a country to have its sovereignty respected.
No it does not seem like an extraordinary claim, unless this is your first time hearing such claims. For the rest of us, we've heard this every decade and it turns out to be entirely untrue.
Nuclear, Overpopulation, Peak Oil, Y2K Bug, Global Warming.
What do you mean, Europeans would love to voluntarily comply with this because they have been of little relevance until American AI companies started to beg for regulation.
I am sure the Europeans rolling up their sleeves right now and getting to "work"
reply