PrismML Squeezes a 27B-Parameter AI Model Onto an iPhone — And Apple Is Paying Attention

Caltech spinout PrismML released Bonsai 27B, the first 27-billion-parameter model capable of running on a phone. Using 1-bit quantization, the model fits in 3.9GB and runs at practical speeds on an iPhone. Apple is reportedly evaluating the technology.

A Caltech spinout just broke one of the quietest ceilings in AI: fitting a model with 27 billion parameters onto a device that fits in your pocket.

PrismML released Bonsai 27B today, a compressed version of Qwen3.6-27B that uses extreme 1-bit and ternary quantization to slash memory requirements by a factor of 14. The 1-bit variant weighs just 3.9GB — small enough to run on an iPhone, which typically exposes only about 6GB of usable memory to applications. The ternary variant, at 5.9GB, targets laptops and desktops.

The performance numbers are striking: 163 tokens per second on an NVIDIA RTX 5090 in 1-bit mode, and 87 tokens per second on Apple's M5 Max. That is not just "technically runs" territory — it is genuinely usable speed for multi-step reasoning, tool calling, and agentic workflows, all running locally without a cloud round-trip.

The model runs natively on Apple hardware through MLX and on NVIDIA GPUs via CUDA, using custom low-bit kernels built for PrismML's hybrid-attention architecture.

PrismML calls its approach "intelligence density" — prioritizing capability per bit rather than raw parameter count. The research builds on work from Caltech.

The story gets more interesting. PrismML CEO Babak Hassibi told CNBC that Apple is among the companies evaluating the startup's models, measuring speed, energy efficiency, and on-device performance. "[Apple is] really evaluating our technology right now," Hassibi said. Apple did not respond to a request for comment.

9to5Mac reported that the two companies have held meetings about how Apple could use PrismML's compression technology. Whether this leads to a partnership or acquisition remains open, but the signal is clear: the industry is hunting for ways to break AI out of the data center.

Commenters on Hacker News, where the story hit the front page with over 560 points, debated whether on-device models at this scale change the economics of cloud inference. The consensus was cautious optimism: running reasoning-capable models locally, with no latency and no privacy trade-off, is a fundamentally different product from a cloud API call.

Bonsai 27B is available now for download, with weights on Hugging Face and inference code for MLX and CUDA.