Multiverse Computing pushes compressed AI models into the mainstream

With private sector default rates above 9.2% (the highest rate in years), venture capital firm Lux Capital recently advised companies relying on AI to confirm their computing power commitments in writing. As financial instability spills over into the AI ​​supply chain, Lux warned that handshake agreements are not enough.

But there is another option entirely. It means completely eliminating dependence on external computing infrastructure. Small-scale AI models that run directly on users’ own devices, with no data center, cloud provider, or counterparty risk, are becoming good enough to be considered. And multiverse computing is also making waves.

The Spanish startup has traditionally maintained a lower profile than some of its peers, but that is changing as demand for AI efficiency increases. After compressing models from leading AI labs like OpenAI, Meta, DeepSeek, and Mistral AI, we made them more widely available by launching both an app that showcases the power of compressed models and an API portal, a gateway for developers to access and build on these models.

The CompactifAI app shares its name with Multiverse’s quantum-inspired compression technology and is an AI chat tool in the vein of ChatGPT and Mistral’s Le Chat. Ask a question and the model will answer. The difference, the company says, is that Multiverse includes a much smaller model called Gilda that can be run both locally and offline.

For end users, this is a taste of edge AI, with no data leaving the device and no connectivity required. However, there are some caveats. Mobile devices must have sufficient RAM and storage. If not, and many older iPhones don’t, the app reverts to a cloud-based model via the API. Routing between local and cloud processing is handled automatically by a system Multiverse has dubbed Ash-Nazg. The name will ring a bell for Tolkien fans, as it references the inscription on the ring from “The Lord of the Rings.” But when apps are routed to the cloud, they lose a key privacy edge in the process.

These limitations mean CompactifAI is not yet ready for mass customer adoption, but that may not have been the goal. According to Sensor Tower data, the app had fewer than 5,000 downloads last month.

The real target is businesses. Today, Multiverse is launching a self-service API portal that gives developers and enterprises direct access to its compression models (no AWS Marketplace required).

“The CompactifAI API portal (now) gives developers direct access to compression models, giving them the transparency and control they need to run them in production,” CEO Enrique Lizaso said in a statement.

Real-time usage monitoring is one of the key features of APIs, and this is no coincidence. Along with the potential benefits of deploying to the edge, reduced computing costs are one of the main reasons why enterprises consider smaller models as an alternative to large language models (LLMs).

It also helps that smaller models are less restrictive than before. Earlier this week, Mistral updated its small model family with the introduction of the Mistral Small 4. It is said to be simultaneously optimized for common chat, coding, agent tasks, and reasoning. The French company also released Forge, a system that allows companies to build custom models, including small-scale models that allow companies to choose the tradeoffs that are most acceptable for their use case.

Multiverse’s recent results also suggest that the gap with LLM is closing. Its latest compression model, HyperNova 60B 2602, is built on gpt-oss-120b, an OpenAI model whose underlying code is publicly available. The company claims that it now offers faster responses at a lower cost than the original version. This is a particularly important benefit for agentic coding workflows, where AI autonomously completes complex, multi-step programming tasks.

Creating models small enough to work on mobile devices while remaining useful is a major challenge. Apple Intelligence avoided this problem by combining on-device and cloud models. Multiverse’s CompactifAI app can also route requests to gpt-oss-120b via the API, but its primary purpose is to demonstrate that local models like Gilda and future alternatives have benefits beyond cost savings.

For workers in critical fields, models that can be run locally without connecting to the cloud offer greater privacy and resiliency. But the greater value lies in the business use cases this enables. For example, incorporating AI into drones, satellites, and other environments where connectivity is commonplace.

The company already serves more than 100 customers around the world, including Bank of Canada, Bosch and Iberdrola, but expanding its customer base could help it raise even more funding. After raising $215 million in Series B last year, the company is now rumored to be raising a new €500 million funding round at a valuation of more than €1.5 billion.