27B dense model stuffed into a pocket, this compression ratio is a bit outrageous—Apple finally can't sit still?

CoinWorld news, PrismML stated that it has compressed Alibaba's open-source model Qwen3.6-27B to approximately 4GB and successfully run it on iPhone 17 Pro. Qwen3.6-27B is a dense model with 27 billion parameters, where all parameters participate in computation during inference, rather than a MOE model that only activates some parameters. PrismML claims that the model size was reduced from about 54GB to less than 4GB without significant performance degradation, making it suitable for complex conversations, reasoning, agent, and coding tasks. Apple has reached out to PrismML to discuss applying this technology to its devices, although advanced Siri still relies on cloud-based models.