27 Billion Weights, One Bit Each

Apple is reportedly in talks with PrismML, a small startup born at Caltech. The startup has already run Qwen 3.6, a 27-billion-parameter model from Alibaba, entirely on an iPhone 17 Pro, according to The Information. A model that size normally needs 54 GB of memory. The version PrismML demonstrated fits in under 4 GB. For now these are meetings, not a deal.

The biggest model Apple currently runs on-device, AFM 3 Core Advanced, has 20 billion parameters, and only 1 to 4 billion are active at a time. PrismML's model keeps all 27 billion active. That gap is the reason for the meetings.

🔬 Each of the model's 27 billion numbers, its weights, normally takes 16 bits. That is the whole 54 GB. PrismML stores each weight in a single bit, and one bit can only say two things, so a weight becomes just a direction, plus or minus. Its precision is gone. Two things replace it. Small groups of weights share one precisely stored number, the scale, which sets their common size. And the model is trained under this constraint from day one, so billions of coarse weights learn to compensate for each other. PrismML says its models are built this way from the start, not compressed after training.

📦 The technology is already public. PrismML ships 1-bit text models under Apache 2.0, an 8B that takes 1.15 GB instead of 16, and image models built the same way, plus a WebGPU demo running a 1-bit model directly in the browser, nothing to install.

📉 Their 1-bit 8B scores 65.7 out of 100 on MMLU Redux, a broad knowledge test, where a full-precision model of the same size scores 83. Squeezing the weights costs quality, and the price depends on the task. PrismML publishes these numbers themselves.

📈 If the 27B holds up, more of Apple Intelligence can move fully on-device. Local means private by default, no network round-trip, and no server bill. It is also fast, since generating a token means reading all the weights from memory, and fewer bits read faster. Their 8B already runs at about 44 tokens per second on an iPhone 17 Pro Max.

PrismML says the 27B ships as open source on July 14. From that day anyone can check the claim.

The 8B is already out (huggingface.co/prism-ml, Apache 2.0). It runs on a Mac with a few lines of Python.