PrismML Runs 27B Qwen on iPhone as Apple Reportedly Circles

A 27-billion-parameter model reportedly running locally on an iPhone 17 Pro is the kind of claim that reframes how consumer AI gets delivered, and it is why The Information reported that Apple has been in acquisition talks with a small Caltech-founded startup called PrismML.

The specific claim is that PrismML compressed Alibaba's open-weights Qwen 3.6 27B model from roughly 54 GB down to under 4 GB and got it running on a shipping phone, with all 27 billion parameters kept simultaneously active. According to Investing.com's summary of the reporting, the underlying technique is a family of ultra-dense 1-bit and ternary weight architectures that the company says can reduce memory footprint by up to 14x and speed inference up to 8x. Qwen 3.6 27B is itself an interesting base, since Alibaba has claimed benchmark parity with Claude 4.5 Opus on several evaluations.

The strategic angle is why Apple is interested. Meta, Microsoft, and Amazon are reportedly committing hundreds of billions of dollars to build out AI data centers that serve hosted inference. Apple's whole product identity rests on the device in your pocket, and if a small research team can put a frontier-class model on that device, the case for paying an OpenAI-style vendor per request weakens considerably. PrismML has raised about $16.25 million from Khosla Ventures, Cerberus, and Google, the kind of round a strategic acquirer can absorb without noticing.

The honest caveat is that the reporting is single-sourced, benchmark parity claims for aggressively compressed models often do not survive contact with real users on real tasks, and neither the reporting nor its summarizers give latency, battery, or held-out evaluation numbers. What the piece also does not answer is how close the acquisition talks actually are, or whether Apple would keep PrismML's models general or fold them into its own foundation-model roadmap. The direction is the part worth watching: for anyone building AI products, planning as if hosted inference is the only viable delivery path is starting to look like a bet that ages badly.