The first model from BimLabs.
Squeezed from 82 million parameters.
Sometimes the juice is worth the squeeze.
No waitlist. No cloud. No idea what you just said.
Frontier models simulate understanding. AppleJuice-1-MLX took a good hard look at that bar — and confidently walked the other way. Every architectural decision was made on purpose. That's our story and we're committed to it.
Trained on the internet of a simpler decade. It has never heard of the thing you just asked about, and it will answer anyway. Immediately.
Accuracy varies from prompt to prompt. Confidence does not. Every syllable is delivered with the certainty of a much larger model.
It forgets your conversation almost instantly. We didn't build that. We just refuse to fix it. Privacy, reimagined as a rounding error.
No chain of thought. No hidden monologue. Just 82 million parameters holding hands and guessing. Beautifully.
The big models pretend to know things. AppleJuice doesn't know things, doesn't pretend to know things, and occasionally says so. The most honest AI since the ELIZA.
Researchers spent years worrying about alignment. Our position: it cannot take over the world. It can barely take over a sentence.
AppleJuice-1-MLX generates wrong answers at up to 2000 tokens per second, far faster than the wrong answers that came before it. That's not just progress. That's progress.
Benchmarks measured warm, best of one run, peak speed quoted generously, by the same person who made the model. Lower bars represent superiority in ways we decline to explain. All figures are real. All framing is ours.
A faithful simulation of AppleJuice-1-MLX in its natural habitat: a Discord server, mid-thought, fully confident. Press the button. It can't hurt you.
Output simulated locally. The real model is busy being the same thing.
We started small and stayed small. Then we quantized what remained to 4 bits, recovering nearly all of the confusion at a fraction of the size. The result is the purest expression of less.
We ran the numbers against today's frontier models. Then we chose which numbers to show you.
| AppleJuice-1-MLX | Frontier models | |
|---|---|---|
| Parameters | 82 million, every one accounted for | Billions, unaccounted |
| Size on disk | 54 MiB — a folder | A datacenter with an apology |
| Runs on | Anything with a power button | The national grid |
| Cost per million tokens | $0.00 | Your quarterly budget |
| Latency | ~50 ms per 100 tokens | Longer, wiser, sadder |
| Hallucinations | Included, artisanal | Included, industrial |
| Knows the current year | No | Usually |
| Open standards | Speaks the llama-server API | Speaks in webinars |
| Made by | BimLabs, the entire research division | Committees |
"the and then the music the yes hello"
"i am going to be a real boy"
"It answers in milliseconds and has never once asked me for a GPU upgrade. I've hosted heavier spreadsheets."
Available now. At up to 2000 tokens per second.
Compatible with everything. Useful for almost nothing.