PrismML Launches Bonsai 27B, the First 27-Billion-Parameter Model to Run Directly on an iPhone

My take: Bonsai 27B shifts the central argument for local mobile models: until now, the standard for running a model offline on a phone was 7B or 8B parameters. With 27B parameters compressed to just 3.9GB through 1-bit quantization, PrismML demonstrates that a serious-size model can run at 11 tokens per second on an iPhone 17 Pro, retaining 90% of full-precision benchmark quality.

The practical implication is direct: the model runs without an internet connection, without sending data to external servers, with all processing on the device. For professionals in regulated industries, or for those working in contexts where privacy matters, that difference can be decisive. Distributing the weights under an Apache 2.0 license makes it accessible to any technical team starting today.

Before treating the benchmarks as settled, it is worth noting that the numbers come from PrismML itself, which has a direct interest in positioning its launch. Waiting for independent evaluations before relying on those figures is the most reasonable approach.

Could your workflow benefit from processing data with an AI model locally, without your information ever leaving the device?