Breaking

PrismML Says It Shrank a 27-Billion-Parameter Model to Run Fully on an iPhone, Drawing Apple's Interest

July 11, 2026 at 11:01 EDT

Foundation Models
Open Source
Infra & Chips

PrismML, a Caltech spinout backed by Khosla Ventures, claims it has compressed Alibaba's 27-billion-parameter open-source model to under 4GB and run it in full on an iPhone 17 Pro — a scale far beyond the compact models typically deployed on phones today. The startup says it plans to open-source the compressed model on July 14, and Apple has reportedly met with the company to examine its technology.

July 14 · PrismML — Caltech spinout

A 27-billion-parameter model, squeezed under 4GB, running fully on an iPhone

PrismML says it compressed Alibaba's dense Qwen 3.6 from a ~54GB footprint to below 4GB — with every parameter active at once — and ran it on an iPhone 17 Pro. Apple has reportedly met with the company to examine the technology.

The compression claim — drawn to scale

Standard precision

~54 GB

PrismML compressed

< 4 GB

A ~14× smaller footprint — all 27 billion parameters kept simultaneously active on the phone.

14×

lower memory use

8×

faster inference

5×

better energy efficiency

10×+

"intelligence density"

Dense vs. sparse — active parameters on device

Apple on-device (sparse)

1–4B active

PrismML (dense, all active)

27B active

Keeping the full parameter set active is what PrismML says enables advanced reasoning, coding help and autonomous agents locally — no cloud round-trips.

What's already shipped

Bonsai models: 8B, 4B, 1.7B (Apache 2.0)
Custom kernels for Apple's Metal stack
Support for MLX and llama.cpp
$16.25M seed — Khosla, Cerberus, Caltech

Still to be proven

Headline figures are the company's own
Independent, real-world testing is limited
Relies on 1-bit & ternary quantization
Open-source release will invite scrutiny

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.