PrismML Releases Bonsai 27B: 27B-Parameter AI Model Runs on iPhone

What happened: PrismML unveiled Bonsai 27B, a 27-billion-parameter multimodal AI model that can run locally on an iPhone 17 Pro Max (at 11 tokens/sec) and laptops, thanks to aggressive model compression. The 1-bit variant is just 3.9 GB, while the ternary version is 5.9 GB—roughly 14x smaller than full-precision models. PrismML claims its ternary variant retains 94.6% of full-precision performance across 15 benchmarks, though these results are self-reported and not yet independently verified.

Why it matters: Bonsai 27B’s ability to run on consumer hardware could accelerate the shift toward private, on-device AI. The model is based on Alibaba’s Qwen3.6 27B, with PrismML’s innovation focused on compression rather than architecture. Apple is reportedly exploring the technology, but practical deployment remains limited to the latest devices. The release raises questions about the trade-offs between compression and performance in real-world applications.

Source: Decrypt