How Siri, Fueled by Gemini, is Anticipated to Operate Internally
**Google and Apple’s $1 Billion Collaboration: The Future of Siri with Gemini**
Earlier this week, Bloomberg disclosed that Google and Apple are close to finalizing a substantial annual agreement worth $1 billion for a variant of the Gemini model, intended to upgrade Siri’s functionalities next year. This collaboration not only signifies a major financial investment but also introduces a crucial transformation in the framework that will affect user experience.
### Grasping the 1.2 Trillion Parameters
As per Bloomberg’s findings, Google is projected to supply Apple with a model featuring 1.2 trillion parameters, which will function on Apple’s Private Cloud Compute servers. This setup guarantees that Google will not have access to the data processed by this model, a significant benefit for user privacy.
The impressive scale of a 1.2 trillion parameter model is noteworthy, but making direct comparisons with competing models is intricate. Recently, prominent AI labs like OpenAI and Anthropic have stopped revealing the parameter counts for their leading models, resulting in speculation surrounding the actual capabilities of models like GPT-5 and Claude Sonnet 4.5. Estimates fluctuate widely, with some indicating they possess fewer than a trillion parameters, while others suggest they could scale to several trillion.
### The Mixture of Experts Framework
A prevalent characteristic among many sophisticated AI models is the employment of a mixture of experts (MoE) framework. Apple has already integrated a version of MoE in its current cloud-based model, which is rumored to comprise 150 billion parameters.
MoE organizes a model with numerous specialized sub-networks, or ‘experts.’ For each input, only a selection of these experts is activated, improving computational efficiency and speed. This allows models to sustain elevated parameter counts while reducing inference costs, as not every parameter needs to be engaged for each input.
For example, a model with 1.2 trillion parameters may activate merely 2-4 experts per input, implying that around 75-150 billion parameters are employed at any given time. This technique allows the model to operate at a high level while keeping computational requirements manageable.
### Future Consequences for Siri
While there have been no detailed reports outlining the framework of the model Google may deliver to Apple, it is highly probable that the MoE approach will be essential for effective operation, particularly given the model’s size. The performance of the Gemini-enhanced Siri relative to other models available at its launch next year remains to be observed, but the incorporation of such cutting-edge technology could greatly improve user engagement and functionality.
In summary, the partnership between Google and Apple signifies a major advancement in AI technology, particularly in the domain of virtual assistants. The anticipated deployment of a 1.2 trillion parameter model, likely making use of a mixture of experts architecture, could transform how Siri functions and communicates with users, establishing a new benchmark in the industry.
Read More
