The startup Multiverse Computing has focused its objective on addressing the problem of large languages, which Some companies cannot assimilate precisely because of that. They do this thanks to the CompactifAI tool.
This company, considered soonicorn ―which means that it is in an intermediate stage but is on the verge of becoming a unicorn― has developed a language model compression technology that It is completely free.
Its open source artificial intelligence model is Hypernova 60B, It is designed for companies that cannot afford certain artificial intelligence models and is available for download on the Hugging Face platform.
The company has explained the greatness of this invention in a press release, where it has indicated that this model is a 50% compressed version of OpenAI’s gpt-oss-120B and that it now offers improvements in different sections with respect to the previous version.
A solution to budget limitations
Those responsible for this Basque company have explained what they want to achieve with this tool and why they have proposed making it available to users free of charge.
So, they have done so, among other things, because «developers are increasingly constrained by massive infrastructure requirements, which limits their ability to test, meaningfully refine and implement AI.”
This means that CompactifAI focuses on creating efficient and high-performance models, which reduce these barriers, while preserving the capabilities advanced reasoning.
That way, you can launch versions of the most powerful models in the world completely free of charge, also eliminating the trade-offs and trade-offs between size, accuracy and accessibility.
And how does he do it?, you might ask. You have a very simple way: apply quantum-inspired mathematics to analyze and reorganize neural networks, preserving only their components with the highest information content.
That way, you can reduce model size by up to 95%, maintaining accuracy within 2-3%. Based on what has been achieved to date by other companies, these figures are round. And the industry standard is characterized by a loss of precision of 20-30% with compression techniques of 50-60%.
Enhancements to advanced coding applications
Multiverse Computing has highlighted that HyperNova 60B has demonstrated that it can achieve frontier-level reasoning performance with «considerably lower» memory, power and infrastructure requirements at a fraction of the typical cost.
It also features substantial improvements in key benchmarks such as comparison to the previous version and can work with advanced encoding applications such as Tau2-Bench, Terminal Bench Hard y BFCL v4.
Multiverse has confirmed that it continues to expand access to production-ready AI models, designed for deployment in enterprise, research and public sector environments.
With this, he has specified that he plans to continue launching models and additional open source updates throughout the year, expanding its focus to multiple sizes and use cases, whether enterprise systems or applications.