Multiverse Computing pushes its compressed AI models into the mainstream

In a climate where private company defaults have surged to over 9.2%, a record high in recent years, venture capital firm Lux Capital has urged AI-dependent businesses to secure their computing capacity agreements in writing. With the AI supply chain facing financial uncertainties, Lux cautioned that informal agreements could fall short. However, an alternative approach is emerging: utilizing smaller AI models that operate directly on users' devices, eliminating the need for data centers or cloud providers. Multiverse Computing, a Spanish startup that has previously flown under the radar, is stepping into the spotlight as demand for efficient AI solutions rises. The company has successfully compressed models from leading AI developers, including OpenAI, Meta, DeepSeek, and Mistral AI. It has introduced both an application showcasing its compressed models and an API portal that enables developers to access and build with these technologies more readily. The CompactifAI app, named after Multiverse's innovative quantum-inspired compression technology, functions similarly to popular AI chat tools like ChatGPT. Users can ask questions, and the app provides answers using a compact model called Gilda, which can operate offline on local devices. This feature offers users a glimpse of edge AI, ensuring that data remains on their devices without requiring a constant internet connection. However, the app's performance hinges on the device's RAM and storage capacity, and for those with older models, it will revert to cloud-based processing via API, sacrificing some privacy in the process. The transition between local and cloud processing is managed by a system dubbed Ash Nazg, a nod to the famous inscription from “The Lord of the Rings.” While CompactifAI is not yet primed for widespread consumer adoption—evidenced by fewer than 5,000 downloads last month—the company is targeting businesses. Multiverse has launched a self-serve API portal that allows developers and enterprises to directly access its compressed models without relying on AWS Marketplace. CEO Enrique Lizaso emphasized that this new portal provides transparency and control essential for deploying models in production environments. A standout feature of the API is real-time usage monitoring, underscoring the appeal of smaller models for enterprises seeking lower computing costs as an alternative to large language models (LLMs). Recent advancements have made these small models more capable. Mistral, for instance, recently updated its small model lineup with Mistral Small 4, which is designed for a range of tasks including general chat and coding. Multiverse's latest compressed model, HyperNova 60B 2602, is built on OpenAI's gpt-oss-120b code and reportedly delivers quicker responses at a lower cost than its predecessor. The challenge remains to create models that are compact enough for mobile devices while still providing substantial functionality. Multiverse's goal extends beyond cost efficiency; it aims to demonstrate that local models offer significant privacy and resilience advantages, particularly for professionals in critical fields. The potential applications for this technology are vast, including integration into drones and satellites where reliable internet access is not guaranteed. Currently, Multiverse serves over 100 global clients, including the Bank of Canada, Bosch, and Iberdrola. The company is also rumored to be on the verge of securing an additional €500 million funding round, building on its prior success in raising $215 million in Series B funding last year, all while valuing the company at over €1.5 billion.