Honor Magic V5 will be able to run AI locally on the device

Honor has introduced the industry's first large language model (LLM) that runs directly on the device itself, rather than in the cloud. The manufacturer promises to launch this feature on the international version of Honor Magic V5 later this week.

Why is it useful?

Working directly on the device allows you to get faster speeds and reduce delays. For example, translating a conversation on the fly is only possible when working directly on the device. Also, local AI is more confidential - your data does not physically leave the device.

Details.

Honor technology allows you to reduce the size of language packs - instead of the standard 3-4 GB for six languages (Chinese, English, German, French, Spanish, Italian), language packs have been reduced to 800 MB, freeing up about 2.78 GB of space on the phone.

Voice processing and translation is instantaneous, before the end of the sentence. Honor claims that this allows for 38% faster processing (compared to cloud-based LLMs) and 16% better recognition accuracy. The developments have been confirmed by two scientific papers at INTERSPEECH 2025, including:

• MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition - for fast and accurate speech recognition.
• Novel Parasitic Dual-Scale Modeling for Efficient and Accurate Multilingual Speech Translation - to accelerate the processing of large models with multilingual support.

It is likely that artificial intelligence technologies will continue to develop, including on local devices. A few months ago, Google already installed LLM on the Pixel 6 smartphone, but it was not a 6-language model, but a stripped-down one adapted to recognise the "language" of dolphins.