Multiverse Computing: Compressed AI Revolutionizes Access

Lux Capital Warns on AI Commitments

In light of a tense economic situation, Lux Capital, an influential venture capital firm, has recently issued a warning to companies relying on artificial intelligence. With the failure rate of private companies reaching a concerning 9.2%, Lux Capital recommends formalizing all commitments related to computing capacity in writing. This precaution aims to address the financial instability currently shaking the AI supply chain, rendering verbal agreements insufficient.

Towards Independence from Cloud Infrastructures

In this context, an alternative is emerging: abandoning dependence on external computing infrastructures. Smaller AI models, capable of running directly on users' devices without the need for data centers or cloud providers, are gaining performance and popularity. Multiverse Computing, a Spanish startup, is part of this trend by developing solutions that mitigate counterparty risks.

Multiverse Computing and Its Ambitions

Until recently, Multiverse Computing operated quietly in the tech sector. However, the growing demand for efficient AI solutions has propelled the company into the spotlight. By compressing models from major labs such as OpenAI, Meta, DeepSeek, and Mistral AI, Multiverse has launched an application and an API portal. These tools facilitate developers' access to compressed models, thereby expanding their availability.

CompactifAI: An Innovative Application

The CompactifAI application, named after the compression technology inspired by quantum mechanics developed by Multiverse, presents itself as an AI chat tool. Comparable to ChatGPT or Mistral's Chat, it allows users to ask questions to which the model responds. The uniqueness of CompactifAI lies in the integration of Gilda, a model compact enough to operate locally and offline, according to the company.

Technical Challenges of CompactifAI

For users, CompactifAI offers a glimpse into edge AI, where data remains on the device without requiring an internet connection. However, this technology imposes hardware requirements: devices must have sufficient RAM and storage. If these conditions are not met, as is the case for many older iPhone models, the application switches to cloud models via the API. This routing, managed by a system named Ash Nazg, inspired by Tolkien's universe, compromises the privacy advantage when the cloud is accessed.

Limited but Strategic Adoption

These constraints explain why CompactifAI is not yet ready for mass adoption by the general public, although this may not be Multiverse's primary goal. According to Sensor Tower, the application was downloaded fewer than 5,000 times last month. The company's real objective seems to be attracting businesses by launching a self-service API portal that offers direct access to its compressed models without going through AWS Marketplace.

An API Portal for Businesses

Multiverse's CEO, Enrique Lizaso, stated that the CompactifAI API portal allows developers to access compressed models directly, with the transparency and control necessary for production deployment. One of the key features of the API is real-time usage monitoring, which is significant. In addition to the potential benefits of edge deployment, reducing computing costs is a major reason why companies are considering smaller models as an alternative to large language models (LLMs).

Evolution of Small AI Models

Small AI models have evolved and are less limited than before. Recently, Mistral updated its lineup with Mistral Small 4, optimized for chat, programming, agentic tasks, and reasoning. The company also launched Forge, a system allowing businesses to create custom models, including small models tailored to their specific needs.

Advances from Multiverse Computing

Recent results from Multiverse show that the gap with LLMs is narrowing. Its latest compressed model, HyperNova 60B 2602, based on OpenAI's gpt-oss-120b, offers faster responses at a lower cost. This advantage is crucial for agentic programming workflows, where AI autonomously accomplishes complex, multi-step tasks.

Challenges and Opportunities of Local Models

Making models small enough to run on mobile devices while remaining performant is a significant challenge. Apple Intelligence has circumvented this issue by combining an on-device model with a cloud model. CompactifAI can also route queries to gpt-oss-120b via the API, but its primary goal is to demonstrate that local models like Gilda offer benefits beyond cost savings.

Outlook for Businesses and Critical Sectors

For professionals in critical sectors, a model operating locally without a cloud connection offers greater privacy and resilience. The real added value lies in the business opportunities this can unlock, such as integrating AI into drones, satellites, and other environments where connectivity is uncertain.

Expansion and Funding of Multiverse

Multiverse Computing already has over 100 clients worldwide, including the Bank of Canada, Bosch, and Iberdrola. Expanding its client base could enable it to access more funding. After raising $215 million in a Series B funding round last year, rumors are circulating about a new funding round of €500 million, valuing the company at over €1.5 billion.