Meta Platforms Inc. plans to release a mobile version of its Meta AI artificial intelligence assistant.

The app is expected to debut in the second quarter, CNBC reported late Thursday.

Meta AI is a ChatGPT-like chatbot service that debuted in September 2023. It’s currently available in Facebook, Instagram, WhatsApp and Messenger as well as through a standalone website. The chatbot can perform many of the same tasks as ChatGPT’s consumer-focused versions.

Alongside the standalone mobile app, Meta reportedly plans to test a paid Meta AI subscription. It’s unclear whether the offering will launch at the same time as the app. There’s no word on what additional features the subscription might provide.

Meta’s reported plan to launch an app for its chatbot doesn’t come as a surprise. In January, a Threads user suggested that the company should develop such an app to make the chatbot more useful. Meta Chief Executive Officer Mark Zuckerberg responded with an emoji expressing agreement.

Besides suggesting a standalone app, last month’s Threads post also listed a number of features that Meta AI’s upcoming mobile version should include. One of the items on the list was support for “lightweight offline use.” Meta has already developed the core technology necessary to power an offline mobile chatbot.

Many AI apps are powered by a large language model hosted in the cloud, which means they require a network connection to access that model. To work offline, a chatbot app must include an LLM lightweight enough to run on the user’s device. In September, Meta released two lightweight models called Llama 3.2 1B and Llama 3.2 3B that meet this requirement. 

Smartphones are powered by systems-on-chip, processors that combine multiple compute modules in a single package. Those modules usually include a central processing unit based on Arm Holdings plc blueprints. When Llama 3.2 1B and Llama 3.2 3B made their debut, Meta revealed that it optimized them to run on Arm silicon. That means the company could potentially use them to power its upcoming Meta AI app. 

A few weeks after introducing the models, Meta rolled out upgraded versions that run up to four times faster using 41% less memory. The efficiency gains were achieved with a machine learning method called quantization. It works by compressing data before entering it into a language model, which boosts response speeds at the expense of lower output quality. 

For Llama 3.2 1B and Llama 3.2 3B, Meta developed a custom quantization algorithm dubbed SpinQuant. It doesn’t decrease models’ output quality as much as earlier implementations. Additionally, the company worked with Qualcomm Inc. and MediaTek Inc. to optimize the quantized Llama 3. models for their mobile SoCs.

It’s unclear what features will be available in the paid subscription that Meta is expected to roll out alongside the Meta AI app. If competing chatbot apps are any indication, the plan might offer higher usage caps or access to more advanced LLMs than the free version. One of those premium LLMs might be Meta’s upcoming Llama 4 algorithm, which is expected to roll out later this year.

Image: Meta