In a landscape where private company defaults are at a staggering 9.2%, VC firm Lux Capital has recently urged AI-dependent companies to secure their compute capacity commitments in writing. This recommendation comes amidst growing financial instability affecting the AI supply chain, highlighting that informal agreements may no longer suffice.
However, there's a transformative alternative: the emergence of smaller AI models that operate directly on users' devices, eliminating reliance on external compute infrastructure. Spanish startup Multiverse Computing is stepping into this arena, offering innovative solutions that promise to redefine the AI experience.
Although Multiverse has maintained a relatively low profile compared to some competitors, the increasing demand for AI efficiency is catalyzing change. The company has successfully compressed models from notable AI labs such as OpenAI and Meta, launching the CompactifAI app and an API portal that allows developers to access and utilize these models more broadly.
The CompactifAI app, named after Multiverse's quantum-inspired compression technology, serves as an AI chat tool similar to ChatGPT. Users can engage with the app, which features the Gilda model--compact enough to function locally and offline. This innovation enables users to leverage AI while keeping their data private, free from cloud dependency.
Nonetheless, users must ensure their devices possess sufficient RAM and storage; otherwise, the app defaults to cloud-based models via API. This seamless transition is managed by a system called Ash Nazg, a nod to Tolkien's lore. However, this cloud reliance diminishes the app's primary privacy advantage.
Currently, the CompactifAI app is not yet poised for widespread consumer adoption, with fewer than 5,000 downloads reported last month. Instead, Multiverse is targeting businesses with the launch of its self-serve API portal, providing direct access to compressed models without the need for AWS Marketplace.
CEO Enrique Lizaso emphasized that this API portal offers developers the transparency and control necessary for production environments. One of its key features is real-time usage monitoring, which aligns with the growing interest from enterprises in deploying smaller models as cost-effective alternatives to large language models (LLMs).
The landscape for small models is evolving, as demonstrated by Mistral's recent updates to its model family, optimizing it for various tasks. Multiverse's advancements suggest that the gap between its compressed models and traditional LLMs is narrowing, with its latest model, HyperNova 60B 2602, claiming to deliver faster responses at a lower cost.
Creating models that can efficiently run on mobile devices presents significant challenges. Yet, the advantages of local processing--particularly in terms of privacy and resilience--are substantial. This technology could revolutionize applications in drones and satellites, where connectivity is not guaranteed.
Currently serving over 100 global customers, including the Bank of Canada and Bosch, Multiverse is also rumored to be raising €500 million in funding, potentially elevating its valuation to over €1.5 billion. The future of AI could very well be shaped by these innovative, compact solutions.