A new local version of personal assistant service has been released, which is the local version of Muse -Product form: Supports automated tasks such as email, schedule, meeting minutes, dictation, and private iMessage, with plans to cover Mac, iOS, Windows, Linux, Android, and NVIDIA. Support privacy, no ads, no token fees, etc. -Based on Qwen 3.8 27B open-source weight model: Their Underdog 27B is based on Alibaba Qwen 3.8 27B. The post compares the Claude Opus 4.6 Max with public benchmarks from six months ago, claiming to have surpassed it in many tasks and can run on Mac or iPhone. Daily tasks (finding flight confirmation letters, claims materials, reservations, emails, schedules) do not require a trillion parameter model -Privacy feature: The founder believes that data should not be uploaded to the cloud. He cited the email application leak and OpenAI/Anthropic related security incidents when he was 18 years old, and proposed the "AI Y3K": models can identify vulnerabilities faster than humans, so "anything that can be hacked will be hacked". The only data that will not leak from the server is data that has never been uploaded. Underdog claims that the model, context, emails/notes, etc. are all locally encrypted and do not have a server for storing personal data. -The speed relies on a dedicated inference engine. Their Husky (self proclaimed Model Specific Inference) is only written for a specific model and is faster than MLX on Apple chips, up to about 4.5 times faster; Their small model Woof can decode up to approximately 730 tok/s on the M5 Max. I am still working on NVIDIA CUDA Megakernel. And propose 'Underdog's Law': Today's cutting-edge intelligence will land on your device within six months. -Capital financing: support from a16z and others