European renowned artificial intelligence startup Multiverse Computing has recently launched two ultra-miniature AI models, named "ChickBrain" and "SuperFly". The company claims that these may be the smallest AI models in the world with still strong performance, capable of running offline on smartphones, tablets, PCs, and even Internet of Things devices.
The founder Román Orús said that these models are based on the company's self-developed CompactifAI compression technology, which significantly reduces the size of AI models without sacrificing performance. This technology is inspired by quantum physics, different from traditional computer science or machine learning compression methods, and achieves higher compression accuracy. In June this year, Multiverse Computing completed a 189 million euros (about 215 million US dollars) funding, and since its establishment in 2019, it has accumulated about 250 million US dollars in funding.
Details of the two new products:
SuperFly: Compressed from Hugging Face SmolLM2-135, the number of parameters was reduced from 135 million to 94 million, equivalent to the size of a fly's brain, suitable for low computing power scenarios such as home appliances, and can achieve voice commands and simple conversations, for example, starting the washing machine mode through Arduino.
ChickBrain: Compressed from Meta Llama3.18B, the number of parameters is 3.2 billion, which can run locally on a MacBook and slightly outperforms the original version in multiple benchmark tests such as MMLU-Pro, Math500, GSM8K, GPQA Diamond.
Multiverse refers to this series as "Model Zoo", named after the size of animal brains. The company says its goal is not to challenge the performance of top large models, but to retain high practical functions at extremely small sizes.
Currently, Multiverse is in discussions with manufacturers such as Apple, Samsung, Sony, and HP, with HP also being an investor. In addition to directly supplying hardware manufacturers, the company also provides compressed model APIs on AWS, allowing developers to call them at costs lower than the industry average. Furthermore, its compression technology is also used in fields such as image recognition, with clients including BASF, Ally, Moody's, and Bosch.