Multiverse Computing opens a public API with compressed models from OpenAI, Meta and DeepSeek

Multiverse Computing, a Spanish startup in quantum computing and AI optimization, has announced a public API and a demo app for its compressed language models. The catalog already includes compressed versions of flagship products from four major labs: OpenAI, Meta, DeepSeek and Mistral AI. It is the company's first step toward the broader market after several years of mostly enterprise sales.

Multiverse Computing was founded in 2019 in San Sebastián, Spain. The company originally specialized in applying quantum algorithms to optimization problems in the financial sector: insurance calculations, risk management and portfolio optimization. Over time it became clear that mathematical methods from the quantum computing toolkit also apply to optimizing neural networks. That observation led to a new product line: compressing large language models while retaining as much of their capability as possible.

Technically, compressing language models is not a single method but a whole class of approaches. Quantization converts a network's weights from float32 to int8 or int4, cutting memory use by 2-4 times with minimal accuracy loss on most tasks. Pruning removes the least significant neurons and entire layers, reducing computational load. Knowledge distillation transfers the behavior of a large model into a smaller, purpose-trained architecture. Together, these techniques can shrink a model by 4-8 times while keeping 90-95% of its quality on typical tasks.

For business, this opens fundamentally new possibilities: running full LLM inference on devices without expensive GPU clusters, such as industrial controllers, medical equipment, corporate servers and mobile devices. In regulated industries (healthcare, finance, the public sector) this is critical: sensitive data stays inside the organization's perimeter and is not sent to external cloud APIs.

The public API fundamentally changes Multiverse Computing's business model. Until now, the company worked mainly through direct enterprise sales: long negotiations, pilot projects and custom integrations with specific customers. An open API moves the product to a self-service model: a developer can sign up and start working without contacting a sales department. The demo app works as a product showroom: prospective customers can evaluate the quality of the compressed models on real tasks before deciding whether to buy or integrate.

The inclusion of DeepSeek, the Chinese open-source model that became one of the major technology events of early 2026, deserves particular attention. Interest in it is huge, but many Western companies and government bodies worry about depending on Chinese infrastructure when using the official API. Deploying a compressed version locally through Multiverse Computing removes that problem: the model runs entirely inside the corporate perimeter.

Having four different providers in the catalog also signals the company's ambitions: it positions itself not as yet another AI lab but as a provider of agnostic infrastructure that works on top of the whole ecosystem of large models.

Multiverse Computing is betting on a simple idea: cloud AI services meet obvious mass demand but leave a huge niche unserved, namely companies and entire industries that cannot or do not want to depend on external APIs for regulatory, technical or economic reasons. A public catalog of compressed models lowers the barrier to entry into this segment to almost zero. If the technology shows sufficient quality in real production scenarios, the company has a real chance to take a strategically important position in the infrastructure layer of the next-generation AI market.

*Meta is designated an extremist organization and banned in Russia.