Multiverse Computing brings compressed AI models to the mainstream market.

Rising Defaults and the Shift Towards Local AI Models

As private company defaults climb to more than 9.2%—the peak in years—venture capital firm Lux Capital has issued a warning to those relying on artificial intelligence (AI) technologies. They suggest that firms secure their compute capacity commitments in writing rather than relying on informal agreements. With financial uncertainty impacting the AI supply chain, a mere handshake is no longer sufficient.

In this climate of instability, an alternative approach is emerging: leveraging smaller AI models that operate directly on users’ devices. This method eliminates dependency on external compute infrastructure, such as data centers and cloud providers, thereby minimizing counterparty risk. Multiverse Computing, a Spanish startup, is at the forefront of this technological evolution.

Multiverse Computing: A Rising Contender in AI

While Multiverse has previously maintained a lower profile compared to its competitors, the escalating demand for AI efficiency has spurred its visibility. The company specializes in compressing AI models from major laboratories, including OpenAI, Meta, DeepSeek, and Mistral AI. It has rolled out two significant products: an app that highlights the capabilities of its compressed models and an API portal that enables developers to access these models more easily.

The CompactifAI app, named after Multiverse’s innovative quantum-inspired compression technology, functions similarly to AI chatbots like ChatGPT and Mistral’s Le Chat. It allows users to ask questions and receive answers. What sets CompactifAI apart is its integration of Gilda, a model sufficiently small to operate locally and offline, ensuring that user data remains private and secure.

Advantages of AI on the Edge

For end-users, CompactifAI offers a glimpse of AI functioning at the edge, where data never leaves their devices and does not require internet connectivity. However, this capability comes with a condition: users must possess mobile devices equipped with adequate RAM and storage. In cases where devices fall short—such as many older iPhones—the app automatically defaults to cloud-based models through an API. The routing of requests between local and cloud processing is managed by a system called Ash Nazg, named after the One Ring inscription in J.R.R. Tolkien’s “The Lord of the Rings.” While this system offers flexibility, switching to the cloud compromises the app’s primary privacy advantage.

These limitations indicate that CompactifAI may not yet be prepared for mass-market adoption; Sensor Tower data reveals the app had fewer than 5,000 downloads in the past month. However, the primary target seems to be businesses rather than individual consumers. Multiverse is launching a self-serve API portal that provides direct access to its compressed models—no AWS Marketplace required.

Key Features of the CompactifAI API

The new CompactifAI API portal gives developers and enterprises control and transparency to utilize compressed models in production. CEO Enrique Lizaso emphasized that real-time usage monitoring is a crucial feature, as it helps businesses optimize their operations. Alongside the benefits of edge computing, lower compute costs are a significant reason why enterprises are gravitating toward smaller models as alternatives to large language models (LLMs).

Recent advancements in small models have enhanced their capabilities. For instance, Mistral recently debuted the Mistral Small 4, designed for various functions such as general chatting, coding, agentic tasks, and reasoning. Mistral also introduced Forge, a system that allows businesses to create custom models tailored to their specific needs.

Narrowing the Gap with Large Language Models

Multiverse’s progress suggests that its models may be closing the gap with LLMs. The latest offering, HyperNova 60B 2602, is derived from the public OpenAI gpt-oss-120b model. Multiverse claims its version delivers quicker responses and lower costs compared to its original source, a vital advantage for applications that require complex, multi-step programming tasks.

Creating models that are both compact enough to run on mobile devices and sufficiently functional poses a significant hurdle. Major companies like Apple have tackled this challenge by integrating both on-device and cloud models. While CompactifAI can route requests to the gpt-oss-120b via API, its primary mission is to prove that local models like Gilda—and its upcoming iterations—offer benefits extending beyond mere cost efficiency.

The Importance of Local Processing in Critical Fields

For professionals in critical sectors, having a model that can operate locally without reliance on the cloud enhances privacy and resilience. This capability can be particularly advantageous in high-stakes environments, such as in the deployment of AI in drones, satellites, and other platforms where consistent internet connectivity cannot be ensured.

Multiverse currently serves over 100 global customers, including reputable names like the Bank of Canada, Bosch, and Iberdrola. However, broadening its customer base is essential for unlocking additional funding avenues. Following a successful $215 million Series B funding round last year, there are ongoing discussions about raising another €500 million at a valuation exceeding €1.5 billion.

Conclusion

As private company defaults rise, the urgency for more reliable AI solutions intensifies. Multiverse Computing is positioning itself as a front-runner in this new landscape by allowing businesses to embrace smaller, local models that reduce risks and enhance privacy. With its focus on real-time monitoring and efficiency, Multiverse not only aims to attract enterprise clients but also to be a cornerstone in the evolution of AI technology that operates directly on users’ devices.

As the demand for scalable, cost-effective, and secure AI solutions grows, Multiverse stands poised to redefine the future of artificial intelligence, offering a promising alternative to traditional computing methods.

Thanks for reading. Please let us know your thoughts and ideas in the comment section down below.

Source link
#Multiverse #Computing #pushes #compressed #models #mainstream

About The Author

Emmanuel Kesse

See author's posts