A language model with 27 billion parameters — all fully active, no sparse shortcuts — running on an iPhone. That’s the promise from PrismML, a startup born from mathematical research at the California Institute of Technology and backed by Khosla Ventures, which claims to have miniaturized Qwen 3.6, the open-source LLM developed by Chinese giant Alibaba, shrinking it from roughly 54GB to under 4GB. The public release is set for next Tuesday.

The numbers are striking: a compression factor over ten times, achieved — according to CEO Babak Hassibi — through a “mathematical trick” that doesn’t degrade performance. If confirmed by independent testing, the result topples one of the firmest constraints of on-device inference: the need to trade off capability to fit within a smartphone’s memory and thermal envelope. The model, the startup says, can handle complex chat, reasoning, autonomous agents, and code generation — all functions previously confined to cloud servers or specialized chips with tens of gigabytes of VRAM.

Compression without the compromise

Techniques like quantization, pruning, and distillation are common currency in the LLM world, but the usual trade-off is a loss of quality, especially on structured reasoning tasks. PrismML claims to have avoided that by leveraging algorithms developed in Caltech labs, with patents exclusively licensed to the startup. Detailed technical documentation hasn’t been released, but the model’s open-source availability will let the community verify whether the full spectrum of original performance has been preserved.

Why this is more than a lab experiment

The test on an iPhone 17 Pro isn’t just a symbolic milestone. It signals a potential redrawing of the boundary between what must go to the cloud and what can live in your pocket. Apple’s own on-device models, using a sparse architecture, keep only 1 to 4 billion parameters active at a time out of a 20-billion total. Here, all 27 billion are running simultaneously — a leap in compute density that, if replicable on non-flagship devices, radically changes the economics of inference: zero latency, no data-transmission costs, full user sovereignty over inputs.

Hassibi’s bet is that within three years, 95% of everyday AI will run locally. “You’ll only need to go to the cloud for the top 5% of high-end stuff,” he said, stressing that this forced migration to the device undercuts AI-as-a-Service providers at the root and shifts hardware investment toward ever more powerful consumer chips and NPUs. It’s not a lone vision: the whole movement around small language models and compression techniques is fueled by the assumption that cloud inference is an unacceptable bottleneck for both cost and privacy.

PrismML plans to apply the same method to trillion-parameter models, bringing them into what we now call edge territory. Such a leap would challenge even the most equipped data centers, because it would tilt the “where to run the model” calculus toward local hardware with low power draw. Next week, when the code and weights are publicly available, we’ll start to see whether this announcement is the first brick in a paradigm shift or yet another overclaim in a sector that has already burned through plenty of promises.