As private company defaults climb beyond 9.2%—reaching their highest level in years—venture capital firm Lux Capital has recently urged AI-dependent companies to secure written confirmations for their compute capacity agreements. Given the financial uncertainty spreading through the AI supply chain, Lux emphasized that informal agreements are no longer sufficient.

However, a completely different approach is gaining traction: moving away from external compute infrastructure entirely. Compact AI models that operate directly on a user’s device—eliminating the need for data centers, cloud providers, and counterparty risk—are now reaching a level of capability that makes them a viable option. Among those stepping forward is Multiverse Computing.

While the Spanish startup has maintained a quieter presence compared to some competitors, rising demand for efficient AI is shifting that dynamic. Having successfully compressed models from leading AI labs such as OpenAI, Meta, DeepSeek, and Mistral AI, the company has introduced both an application to demonstrate its compressed models and an API portal. This gateway allows developers to access and build with these models, broadening their availability.

The CompactifAI app, named after Multiverse’s quantum-inspired compression technology, functions as an AI chat tool similar to ChatGPT or Mistral’s Le Chat. Users pose questions, and the model responds. What sets it apart is the integration of Gilda, a model described by the company as sufficiently compact to run locally and offline without an internet connection.

For end users, this offers a glimpse of edge AI, where data remains on their devices and no connection is required. There is an important limitation, however: the mobile device must have adequate RAM and storage. If it doesn’t—which is the case for many older iPhones—the app automatically reverts to cloud-based models via API.

This switching between local and cloud processing is managed by a system Multiverse calls Ash Nazg, a name familiar to Tolkien fans as it references the One Ring inscription from *The Lord of the Rings*. When the app routes queries to the cloud, though, it sacrifices its primary privacy advantage.

These constraints mean CompactifAI isn’t yet positioned for widespread consumer adoption—which may not have been the objective in the first place. According to Sensor Tower data, the app recorded fewer than 5,000 downloads over the past month. The real focus appears to be on business clients.

Today, Multiverse is launching a self-serve API portal that provides developers and enterprises direct access to its compressed models, bypassing platforms like AWS Marketplace. CEO Enrique Lizaso stated, “The CompactifAI API portal now gives developers direct access to compressed models with the transparency and control needed to run them in production.” Real-time usage monitoring is a central feature of the API, a deliberate inclusion given that reduced compute costs are a key incentive for enterprises exploring smaller models as alternatives to large language models.

It also helps that small models are becoming more capable. Earlier this week, Mistral updated its small model lineup with Mistral Small 4, which it describes as optimized for general chat, coding, agentic tasks, and reasoning. The French firm also released Forge, a system enabling enterprises to develop custom models—including compact ones—where they can select trade-offs best suited to their use cases.

Multiverse’s recent achievements indicate the performance gap with LLMs is closing. Its latest compressed model, HyperNova 60B 2602, is based on gpt-oss-120b, an OpenAI model with publicly available underlying code. The company asserts it now delivers faster responses at a lower cost than the original, an especially valuable advantage in agentic coding workflows where AI autonomously handles complex, multi-step programming tasks.

Creating models small enough for mobile devices while retaining usefulness is a significant technical hurdle. Apple Intelligence addressed this by blending an on-device model with a cloud model. Multiverse’s CompactifAI app can also route requests to gpt-oss-120b via API, but its core aim is to demonstrate that local models like Gilda—and its future successors—offer benefits beyond cost savings.

For professionals in sensitive fields, a model that runs locally without cloud connectivity provides enhanced privacy and operational resilience. The greater potential, however, lies in the business applications it enables—such as integrating AI into drones, satellites, and other environments where reliable connectivity cannot be assumed.

The company already serves over 100 global clients, including the Bank of Canada, Bosch, and Iberdrola. Expanding its customer base could support further fundraising efforts. Following a $215 million Series B round last year, rumors suggest Multiverse is now seeking approximately €500 million in new funding at a valuation exceeding €1.5 billion.