I do understand, at least in principle, how marketing works. And thanks to Sam Altman and his disciples, I also understand extremely well how Silicon Valley marketing lies work. On YouTube and X they promise you the moon and stars. “The 1-bit revolution!”, “Low latency!”, “Runs on a toaster!”. Well, congratulations. Maybe it does run extremely fast and barely touches the VRAM on my RTX 4070 - but what use is an answer in milliseconds if it has the intellectual depth of a dried-up Irish puddle?
Why the 1-bit hype sounded exciting at first
But what exactly is all this hype around these models supposed to be? At first I was genuinely curious. Microsoft’s BitNet research actually opened a very interesting door: the idea of throwing all those huge, memory-hungry floating-point weights overboard and restricting them to three simple states - minus one, zero and plus one - is an absolutely brilliant approach. It is easier on the hardware, drastically reduces energy consumption, and I sincerely hope Microsoft continues to push this research forward consistently, because this could genuinely be the future of fast local AI.
And then PrismML showed up with its Bonsai project and apparently thought it would be a wonderful idea to take that forward-looking concept and push it to the absolute extreme. A model trained from the ground up with a ruthless 1-bit architecture - where every single weight is brutally compressed into a pure binary decision, without the neutral zero that Microsoft’s template still had as a stabilising middle ground. For once, this actually sounded like a real technical breakthrough and not just the usual open-source label fraud where some exhausted hobbyist rips a few layers out of a stable Qwen model at three in the morning and then gives it a click-friendly name like “Deckard-Heretic.”
So yes, I genuinely expected a lot from this radical interpretation of the architecture. But, well, I was massively disappointed.
Fast, tiny, and still astonishingly dumb
Half of YouTube is celebrating these things, and of course we are still at the absolute beginning of 1-bit research. But apart from the fact that the dreaded endless loops you regularly get with classical Q1 quantisations do indeed seem to be missing here, I see almost no meaningful difference in the abysmal quality of the answers. Quite the opposite. The moment you ask these supposed miracle models for even the tiniest spark of genuine logic or spatial reasoning, they implode in a way that is almost impressively stupid.
And we are not talking about small, charming AI slip-ups here. We are talking about models that respond to a simple, friendly “Hi, how are you?” by telling me three times in a row what model they are and where they were programmed.
Then I asked the model that one question the voice models from all the major providers are currently failing at. I already wrote a separate article about those voice models here. The question was whether I should walk to a car wash that is only two minutes away on foot - or drive there with the car. See for yourself:
When confronted with the logic error, the model backed off - only to then recommend once again... walking there.
Compare that with 4B models in Q5 or Q6 - none of those models fail quite this spectacularly. And as for speed: the little Bonsai runs on my RTX 4070, and here are the numbers:
And honestly? With a 9B model at Q8 I can easily get 60 to 70 tokens per second. So 91.93 t/s does not exactly knock me off my chair.
Even a simple website already exposes the limits
Then I thought I would at least give the model a chance to build a simple website.
I have linked that example page directly here. Once again, another 8B model from the Qwen family in a better quantisation can throw together a website in no time that does not look like a retro relic from the 1990s.
My verdict on Bonsai, PrismML and 1-bit models
All in all, AI - and local AI in particular - is a field that fascinates me, and one in which we will see massive progress over the next few years. But if I may ask the YouTube clowns for one small favour: please stop hyping models that provide absolutely no practical value. Yes, it gets clicks, because people whose mental altitude matches these models enjoy watching this stuff. That is 2026. Sadly.
Microsoft - please, please keep going with your BitNet research. We need usable open-source models built around exactly this kind of architecture. If we can get models with 20 to 30 billion parameters running cleanly on BitNet, then we will have taken a very big step forward.
It is slowly becoming clear to me that I use this blog mostly to complain. But while a lot of genuinely impressive things are happening in AI, there is also a great deal of garbage. Next I will probably end up writing about the long-awaited, equally disappointing, and yet still somehow hyped Gemma 4. But first, feel free to tell me in the comments what you think about these little Bonsai models and the hype around them.