PrismML crushed a 27 billion parameter AI model down to under 4 gigabytes, small enough to run natively on an iPhone, and Apple is already evaluating the technology.
A Caltech spinout just squeezed a 27 billion parameter AI model down to under 4 gigabytes. Small enough to run on your iPhone. And Apple is already looking at it.
Bonsai 27B landed on July 14, 2026. If you've spent any time around large language models, you know why that matters. PrismML, the startup behind it, took Alibaba's Qwen3.6-27B, a model that normally needs about 54 gigabytes of memory, and crushed it down to 3.9 gigabytes without gutting what it can actually do.
Here's how. PrismML quantized the model to 1-bit and ternary precision, so most of its weights collapse to just three possible values: -1, 0, or +1. That's extreme compression. Each group of weights keeps its own scaling factor, though, stored at full FP16 precision, and that's what saves it from turning to mush: roughly 1.71 bits per weight for the ternary version, applied end to end, across embeddings, attention layers, the MLPs and the language modeling head. Vision components got handled separately, at 4-bit. The result runs on an iPhone 17 Pro at 11 tokens per second, according to PrismML's own published benchmarks. On a Mac with an M5 Max chip, it hits 87 tokens per second.
None of that matters if the model gets dumber in the process. It mostly doesn't. The ternary build retains 94.6% of the full-precision baseline's performance, and the 1-bit version holds onto 89.5%, according to benchmarks reported by MarkTechPost. Math and coding barely move. Tool calling stays within a few points of the original. PrismML released both versions on Hugging Face, alongside a 262,000-token context window inherited from Qwen3.6's hybrid-attention backbone.
That's the release. The bigger story is who's paying attention to it.
Apple Is Already Testing It #
PrismML CEO Babak Hassibi told CNBC that Apple is evaluating the technology, testing its speed, energy use and accuracy alongside other companies interested in the same problem. He was careful to frame it as early. Things are progressing nicely, he said, but he wouldn't say where the talks are headed.
You can guess why Apple is interested anyway. Siri has lagged behind Google Assistant and ChatGPT for years, and Apple's whole pitch to customers rests on privacy and speed, both of which get harder to deliver when a request has to leave the phone, hit a data center, and come back. A model that runs natively on an iPhone 15 or newer, per CNBC's reporting, solves that problem directly. No round trip. No cloud bill. No data leaving the device.
What This Means for the Race to Shrink AI #
Frankly, this is the clearest signal yet that the fight for on-device AI won't be won by throwing more compute at bigger data centers. It'll be won by whoever can shrink a frontier-class model without breaking it, and PrismML, a small Khosla Ventures-backed team spun out of Caltech, just did that to a 27-billion-parameter model with commodity hardware as the target.
That should worry the edge AI startups competing for the same real estate. Most of them have raised money on a bet that Big Tech would move slowly on model compression, since none of Apple, Google or Samsung has publicly matched what PrismML just shipped. A well-funded lab could replicate the technique. But PrismML got there first, published it, and put it directly in front of the company most likely to deploy it at scale.
It also raises a harder question for Apple's own AI roadmap. The company has spent the past two years promising an overhauled Siri built on in-house models, with a public timeline that keeps slipping. If a 27-billion-parameter third-party model can already run this well on existing hardware, licensing or acquiring outside technology becomes a lot more tempting than waiting for an internal team to catch up. Hassibi's own admission that talks are early suggests Apple hasn't decided either way.
What's clear is that the ceiling for phone-based AI just moved. A year ago, running anything close to a 27-billion-parameter model locally, at usable speed, would have sounded like a demo trick. Now it's a downloadable file on Hugging Face, and the company that made it is sitting across the table from Apple.
Also read: Anthropic Bets the Next Trillion Dollar Business Is Installing AI, Not Building It • Rime Raises $24 Million to Put Its AI Voice on Enterprise Phone Lines • Indian Coding Startup Emergent Becomes a Unicorn in Just Over a Year