Multiverse Computing releases CompactifAI API and new compressed AI model

In short: Multiverse Computing says it can shrink popular AI models and is rolling them out through a new model release and a CompactifAI API.

What happened

Multiverse Computing, an AI startup based in Spain, says it has found a way to compress large language models, meaning the text generating AI behind many chatbots. It says its CompactifAI technology can cut a model’s file size by as much as 95%, with only a 2% to 3% drop in accuracy.

On March 2, 2026, the company released HyperNova 60B 2602, a compressed version of OpenAI’s gpt-oss-120B. Multiverse says it reduced the model from 61GB to 32GB, and added improvements for “tool calling” (letting the AI use other software like a helper) and coding agent features (letting it take more steps on its own when writing code). The company made the model available for free on Hugging Face.

Multiverse has also been pushing its compressed models into wider use through its CompactifAI API. An API is like a plug that lets other apps and services connect to Multiverse’s models without building everything themselves. As of March 16, 2026, the company said it added compressed versions of NVIDIA’s Nemotron-3 models, and it already offers compressed models from Meta’s Llama, DeepSeek, and Mistral.

Why it matters

Big AI models can be expensive to run because they need powerful computers and lots of electricity. If Multiverse’s compression works as advertised, it is like taking a huge book and making it into a smaller, easier-to-carry edition that still reads almost the same. That could make advanced AI cheaper to use, and in some cases possible to run on regular computers and even phones.

Source: TechCrunch AI