Multiverse Computing Launches Revolutionary AI Models: Meet "ChickBrain" and "SuperFly"
In an exciting advancement for the world of artificial intelligence, one of Europe’s leading AI startups, Multiverse Computing, has unveiled two groundbreaking AI models named after the brains of a chicken and a fly. These innovative models, humorously dubbed "ChickBrain" and "SuperFly," are not just clever names; they represent a significant leap forward in the field of compact AI technology.
The Smallest High-Performing AI Models
Multiverse Computing has claimed that ChickBrain and SuperFly are among the world’s most compact high-performing AI models, designed to handle a wide array of tasks, including chat, speech, and reasoning. The ability to provide these functionalities in such small packages opens up a realm of possibilities for integration into Internet of Things (IoT) devices and personal devices like smartphones, tablets, and PCs.
On-Device Intelligence
The primary goal behind these miniature models is their ability to function directly on devices, without needing a constant internet connection. As founder Román Orús stated, “We can compress the model so much that they can fit on devices.” This level of on-device intelligence not only enhances user experience but also addresses privacy concerns associated with cloud-based AI processing.
Birth from Quantum Physics
Founded in 2019 and headquartered in Donostia, Spain, Multiverse Computing leverages advanced quantum physics principles in its model compression technology, called “CompactifAI.” The company recently completed a €189 million funding round, bringing its total funding to about $250 million. Orús elaborated on their unique compression algorithm, emphasizing that it transcends typical computer science methods, presenting a more refined approach to AI model size reduction.
A Unique Approach to Model Compression
The compact models released by Multiverse Computing illustrate how a quantum-inspired technique can significantly reduce model size without compromising performance. “It’s a subtle and refined compression algorithm,” Orús described. The startup has already applied this technology to popular models from OpenAI and others, effectively generating compressed versions of open-source AI models, which include Hugging Face’s SmolLM2-135 and Meta’s Llama 3.1.
More Read
Meet the Models: ChickBrain and SuperFly
SuperFly: The Smaller Yet Smart Model
SuperFly is a notably compact model derived from Hugging Face’s SmolLM2-135, which originally features 135 million parameters. With SuperFly’s 94 million parameters, the analogy of its size being comparable to a fly’s brain is both humorous and fitting. This model is specifically crafted to operate on highly restricted datasets, focusing on tasks like voice-controlled commands for home appliances, which can revolutionize the way we interact with everyday products.
ChickBrain: The More Capable Giant
In contrast, ChickBrain boasts 3.2 billion parameters, allowing for enhanced reasoning abilities. Derived from Meta’s Llama 3.1, this model can run on common devices like MacBooks without an internet connection. Notably, ChickBrain achieves competitive performance, even slightly surpassing the original model in various standard benchmarks. Its capabilities extend to language tasks, mathematics, and general knowledge, making it a versatile tool for various applications.
Performance Metrics
Multiverse Computing has shared internal testing results for ChickBrain, demonstrating its effectiveness across multiple benchmarks, including MMLU-Pro, Math 500, and GPQA Diamond. Although the company has not yet released performance metrics for SuperFly, its intended applications are designed around less complex tasks that do not necessitate deep reasoning capabilities.
Collaborations with Major Manufacturers
Multiverse is currently in discussions with leading technology companies, including Apple, Samsung, and Sony, as it aims to integrate its innovative models into mainstream consumer devices. With HP as a key investor, the collaboration potential for implementing these AI models across various platforms is promising.
Broader Applications Beyond Chat AI
The startup is not limited to compressed models for language learning or chat; it also focuses on machine learning applications like image recognition. Its extensive client list includes industry giants like BASF, Ally, and Bosch, indicating a robust demand for its compression technology across diverse sectors.
Accessibility through APIs
For developers looking to integrate these compressed models into their applications, Multiverse offers an API hosted on AWS. This platform allows developers to access their models, often at lower costs compared to competitors, making cutting-edge AI technology more accessible to a wider audience.
With the launch of ChickBrain and SuperFly, Multiverse Computing is redefining the boundaries of what’s possible in AI model performance, shrinking the size while enhancing functionality. The implications for everyday technology are vast, and the potential for practical applications could soon change how we interact with the devices around us.
Inspired by: Source
