Apple eyes PrismML tech that runs 27B AI model on iPhone

Apple is in early talks with startup PrismML about its Bonsai 27B model, a 27‑billion‑parameter LLM compressed to roughly 4 GB and capable of running natively on iPhone, iPad and Mac [1]. The discussions matter because they could let Apple bring server‑class AI capabilities onto its devices without relying on cloud inference, a key differentiator in the mobile AI race.

PrismML’s breakthrough model

PrismML announced Bonsai 27B on July 14, 2026, claiming it runs at up to 163 tokens/second in 1‑bit mode on an NVIDIA RTX 5090 and 87 tokens/second on Apple’s M5 Max chip [1]. The model’s 4 GB footprint fits within the roughly 6 GB of usable memory on a 12 GB iPhone, leaving room for KV cache and activations – a size no conventional 27 B model can achieve [1]. The startup also released the model under an Apache 2.0 license and opened a limited‑time developer preview API [1].

Apple’s interest and market context

The Information reported that Apple has already held meetings with PrismML to explore “ways it could use its technology” [2]. Apple’s broader AI strategy has focused on shrinking large language models for on‑device use, but it has been constrained by the memory demands of server‑sized models [4]. PrismML’s ability to compress the 54 GB Qwen 3.6 model to 4 GB without performance loss directly addresses this bottleneck [4]. If Apple adopts the tech, it could enable more sophisticated on‑device AI features—such as advanced chat, reasoning, and autonomous agents—than the current generation of phone‑friendly models, which typically run only a few billion parameters [2][4].

Competitive implications

Apple’s rivals, notably Google and Microsoft, are investing heavily in cloud‑centric AI services, while Apple has emphasized privacy‑first, on‑device processing. Securing a method to run a 27 B model locally would give Apple a unique value proposition: high‑quality AI without transmitting data to the cloud. The move also aligns with Apple’s recent collaborations, such as co‑developing a new Siri AI with Google for iOS 27 [2]. However, the talks remain “very early,” and no formal partnership has been confirmed [1].

What to watch

• Developer preview uptake – adoption rates of the Bonsai 27B API among Apple developers could signal Apple’s commitment.
• Apple product roadmap – any announcement at WWDC 2026 or later that references on‑device LLMs may indicate integration plans.
• Further meetings – reports of additional Apple‑PrismML discussions or a formal partnership would clarify the trajectory.

If Apple successfully integrates PrismML’s compression tech, it could reshape the mobile AI landscape by delivering server‑grade capabilities on consumer devices, but the outcome hinges on technical validation and strategic alignment in the coming months.