Multiverse AI Unveils Two Tiny Yet Powerful New Models
▼ Summary
– Multiverse Computing released two ultra-small AI models named SuperFly (94M parameters) and ChickBrain (3.2B parameters), designed for IoT devices and local use on smartphones or PCs.
– The models use CompactifAI, a quantum-inspired compression technology that reduces size without sacrificing performance, enabling on-device AI without internet connectivity.
– ChickBrain, a compressed version of Meta’s Llama 3.1 8B, outperforms the original in benchmarks like MMLU-Pro and GSM8K, while SuperFly targets voice-controlled appliances.
– The startup raised €189M in June, backed by investors like HP Tech Ventures, and is in talks with major manufacturers like Apple, Samsung, and Sony for integration.
– Multiverse offers compressed models via an AWS-hosted API with competitive pricing, alongside serving clients like Bosch and Moody’s for machine learning applications.
A European AI startup has unveiled two remarkably compact yet powerful models designed to bring advanced capabilities to everyday devices, without requiring cloud connectivity or significant processing power. Multiverse Computing, known for its expertise in quantum-inspired compression technology, has introduced these models as part of its “Model Zoo,” playfully named after animal brain sizes to reflect their efficiency.
The first, SuperFly, is a streamlined version of Hugging Face’s SmolLM2-135, shrinking its parameters from 135 million to just 94 million, comparable to the neural complexity of a fly’s brain. Despite its diminutive size, SuperFly can process voice commands and troubleshoot issues for IoT devices like smart appliances. Imagine telling your washing machine to “start a quick wash” or asking it why a cycle failed, all without an internet connection.
The second model, ChickBrain, packs a more substantial punch at 3.2 billion parameters, yet remains small enough to operate locally on a MacBook. Derived from Meta’s Llama 3.1 8B, ChickBrain not only matches but slightly outperforms its predecessor in key benchmarks, including language comprehension (MMLU-Pro), mathematical reasoning (Math 500, GSM8K), and general knowledge (GPQA Diamond).
What sets Multiverse apart is its proprietary CompactifAI compression algorithm, which leverages quantum physics principles to reduce model sizes without sacrificing performance. This breakthrough has already attracted major industry players, with the startup securing €189 million in funding earlier this year. Investors include HP Tech Ventures and Toshiba, signaling strong confidence in its potential to revolutionize edge AI.
Beyond consumer electronics, Multiverse’s technology is being adopted by enterprises like Bosch and Moody’s for applications ranging from image recognition to predictive analytics. Developers can also access these compressed models via an AWS-hosted API, often at a lower cost than competing solutions.
While these models won’t rival the largest AI systems in raw power, their ability to deliver high performance in constrained environments makes them a game-changer for embedded systems. As smart devices become ubiquitous, Multiverse’s innovations could redefine how AI integrates into our daily lives, one tiny, efficient brain at a time.
(Source: TechCrunch)