Revolutionizing AI: Tiny Models, Immense Potential
The Genesis of Miniature AI: A \"Model Zoo\" Emerges
Multiverse Computing, a prominent European AI innovator, has launched two diminutive artificial intelligence models, playfully termed the \"Model Zoo.\" These pioneering creations are so compact they have been metaphorically aligned with the brain sizes of a chicken and a fly, yet they deliver exceptional performance in chat, speech, and even complex reasoning tasks.
Empowering Devices with On-Device Intelligence
The newly developed ultra-small models are specifically engineered for integration into Internet of Things (IoT) devices, enabling AI functionalities directly on hardware such as smartphones, tablets, and personal computers, thereby eliminating the necessity for cloud connectivity. This allows for local operation, enhancing data privacy and responsiveness, and expanding the reach of advanced AI capabilities to a multitude of everyday gadgets.
CompactifAI: The Quantum Leap in Model Compression
Multiverse Computing, established in 2019 and headquartered in Donostia, Spain, is at the forefront of AI model compression. The company recently secured a substantial €189 million (approximately $215 million) investment, primarily due to its innovative \"CompactifAI\" technology. This quantum-inspired compression algorithm masterfully shrinks the footprint of AI models without compromising their operational efficiency, setting a new benchmark in the field.
Expanding the Horizon: Broad Compatibility and Superior Efficiency
Multiverse Computing has successfully applied its compression techniques to a diverse range of open-source models, including widely used smaller architectures such as Llama 4 Scout and Mistral Small 3.1. Furthermore, the company has released compressed versions of OpenAI's recently introduced open models and even significantly reduced the size of very large models, such as DeepSeek R1 Slim. Their core focus remains on developing the most compact yet potent AI models available, driving accessibility and efficiency across the AI landscape.
Introducing SuperFly and ChickBrain: Bridging the Gap in AI Accessibility
Multiverse's latest offerings include SuperFly, a highly compressed version of Hugging Face's SmolLM2 135, weighing in at just 94 million parameters—comparable to a fly's brain. This model is engineered for minimal data training, ideal for voice-activated home appliances and troubleshooting. The second model, ChickBrain, a 3.2 billion parameter adaptation of Meta's Llama 3.1 8B, boasts enhanced reasoning capabilities and, remarkably, outperforms its uncompressed counterpart in key benchmarks like MMLU-Pro and Math 500, all while operating efficiently on devices like a MacBook without requiring an internet connection.
Setting New Standards in Performance and Industry Collaboration
While Multiverse's "Model Zoo" may not top leaderboards dominated by massive, state-of-the-art models, their innovative approach ensures that performance remains high despite significant size reduction. The company is actively engaging with leading technology manufacturers, including Apple, Samsung, Sony, and HP, the latter being an investor in their recent funding round. Beyond model sales, Multiverse offers its compressed models through an AWS-hosted API, providing developers with cost-effective and efficient access to their groundbreaking AI solutions, and extending its compression technology to other machine learning applications such as image recognition for major clients like BASF and Bosch.
