Multiverse debuts HyperNova 60B compressed AI model

Multiverse Computing has released HyperNova 60B 2602, a compressed large language model based on OpenAI's gpt-oss-120B, and made it available for free on Hugging Face.

The Spanish firm says the model cuts memory requirements from 61GB to 32GB, which it describes as 50% compression, while keeping near-parity tool-calling performance with the original.

HyperNova 60B 2602 is the latest iteration in Multiverse's HyperNova line. It is positioned as part of a broader push for efficiency-led development of advanced AI models, as infrastructure limits and energy use become central concerns for developers and policymakers.

Compression focus

Model size and hardware demands remain practical constraints for organisations that want to run modern AI systems without relying entirely on large cloud clusters. A model's memory footprint can determine whether it runs on a single server, a workstation, or closer to the edge on constrained devices. Multiverse argues that compression offers an operational alternative to building ever-larger models.

The approach is also tied to European discussions about "sovereign AI," where efficiency becomes more significant as governments and enterprises weigh where models run, what infrastructure they depend on, and how much energy they consume.

Multiverse's compression technology, CompactifAI, uses what it calls quantum-inspired mathematics to analyse and reorganise neural networks. The company says the method retains only the most information-rich components of a model and can reduce model size by as much as 95% while keeping precision within a 2-3% margin.

Multiverse contrasts those figures with what it describes as an industry norm of 20-30% accuracy loss when using techniques that achieve 50-60% compression. No third-party validation was provided with the announcement.

Benchmark claims

HyperNova 60B 2602 includes changes aimed at tool calling and instruction following. Tool calling has become an important capability for developers building agent-like workflows that connect a language model to external functions, code execution, or software tools.

Multiverse reports improvements on several agentic and coding-related benchmarks compared with its prior HyperNova 60B release, including a 5x improvement on Tau2-Bench, a 2x improvement on Terminal Bench Hard, and a 1.5x improvement on BFCL v4.

The company also says the model retains nearly the same tool-calling capability as the full-size OpenAI gpt-oss-120B while running at roughly half the size. It frames the results as evidence that compression can be part of a regular model improvement cycle rather than a one-time optimisation step.

"The launch of HyperNova 60B 2602 demonstrates compression as an iterative process of improvement, not a one-time optimization. Each generation of our compressed models pushes the boundaries of what's possible with efficient AI," said Enrique Lizaso Olmos, CEO of Multiverse Computing.

"By continuing to refine HyperNova and releasing it openly, we're giving developers the tools to experiment, validate, and deploy efficient AI without massive infrastructure investments," Olmos added.

Open distribution

HyperNova 60B 2602 is available through Multiverse's page on Hugging Face, a widely used platform for hosting and distributing machine learning models. Documentation, benchmarks, and integration guides accompany each release, according to the company.

Multiverse says the open release model allows developers, IT teams, and other organisations to evaluate performance and operational fit before wider deployment. It also enables organisations to review security and other considerations during testing.

The release follows an earlier HyperNova 60B model that Multiverse says it launched in January 2026. The company says it incorporated developer feedback into version 2602, including requests for stronger tool-calling performance.

Multiverse says it plans to publish additional compressed models and updates during 2026 across multiple sizes and use cases, including systems designed for edge and device-level deployment. "With this latest model launch, Multiverse continues to expand access to production-ready AI models designed for real-world deployment across enterprise, research, and public sector environments."