Spanish AI company Multiverse Computing has released HyperNova 60B 2602, a compressed version of OpenAI’s gpt-oss-120B, and published it for free on Hugging Face.
The new version cuts the original model’s memory needs from 61GB to 32GB, and Multiverse says it retains near-parity tool-calling performance despite the 50% reduction in size.
In theory, this means a model that once required heavy infrastructure can run on far less hardware. For developers with tighter budgets or energy constraints, that’s a potentially huge advantage.
CompactifAI technology
Multiverse claims gains in agent-focused benchmarks compared to its earlier compressed release. It says HyperNova 60B 2602 delivers a 5x improvement on Tau2-Bench and 2x on Terminal Bench Hard.
Those tests measure tool use and coding workflows rather than simple text replies.
The company’s CompactifAI technology restructures transformer weight matrices using quantum-inspired tensor networks.
Multiverse believes that effective compression offers an alternative to simply building larger and larger models, and links that view to ongoing European discussions around sovereign AI, infrastructure limits, and energy use – so to find out more, I spoke to the company about its compression technology.
- How can you compress an LLM?
Multiverse Computing compresses large language models using its proprietary CompactifAI technology, based on quantum-inspired tensor networks.
Instead of simply removing parameters, CompactifAI restructures the internal weight matrices of transformer models into highly efficient tensor network representations. This mathematical reformulation captures correlations between parameters and eliminates structural redundancy.
The process is applied post-training, meaning the original model does not need to be retrained and no access to the original training data is required.
Using this approach, CompactifAI can reduce memory usage by up to approximately 93% and significantly cut parameter counts, while maintaining strong performance across tasks.
The resulting compressed models are smaller, faster, more energy-efficient, and easier to deploy across cloud, on-premise, and edge environments.
- Can you apply it to every LLM?
It works on transformer-based large language models, including dense foundation models, provided access to the model weights is available.
The technology is architecture-agnostic within the transformer family and does not require changes to the model’s external behavior or APIs.
Compression effectiveness depends on the level of redundancy in the model. Large, overparameterized models typically offer the greatest compression potential.
The primary technical challenge is preserving model accuracy while achieving high compression ratios. This is addressed by carefully controlling tensor decomposition parameters to balance size reduction and performance stability.
Another challenge is ensuring that compressed models maintain robustness across different tasks, including reasoning, multilingual performance, and domain-specific use cases.
Finally, deployment environments vary widely. Compression must be optimized for different hardware targets, latency requirements, and operational constraints.
- What could a good analogy be?
Rewriting the blueprint, not removing bricks: CompactifAI does not simply remove parts of a model. Instead, it rewrites the mathematical blueprint so the same structure is represented more efficiently.
It is like redesigning a building’s internal framework so it uses far less material while preserving strength and functionality.
Another analogy is reorganizing a massive archive into a highly structured system that eliminates duplication. The knowledge remains intact, but it is encoded far more efficiently.
- How do you determine accuracy loss?
Accuracy loss is determined by benchmarking the compressed model against the original on the same tasks and scoring metrics, then measuring the change.
In practice, that includes tool-calling evaluations. Reducing a loss in capability here enables more advanced agentic workflows and coding applications.
- What other companies (perhaps rivals) are working on the same technique
Multiverse Computing’s compression technique is totally unique, based on research into quantum-inspired tensor networks by cofounder and CEO Roman Orus.
Though there are other techniques available for compressing AI models, they come with the trade-off of a much higher degree of accuracy loss.
- Given the fact that LLMs organically evolve over time, what could be the future of your compression (hardware implementation maybe?) or something else?
This compression technique can be applied to upcoming LLMs as well, meaning that in the future, devices such as cars, phones, laptops etc. will be able to run small or nano AI models preinstalled on their hardware.
- Is it hardware agnostic? Does it work better with some hardware (ASIC) than others?
Yes, it’s hardware-agnostic at the model level: CompactifAI compresses the model weights post-training, so the resulting model can be deployed across cloud, on-prem, and edge without changing the model’s external interface.
Inference speedups depend on what was limiting you before: If you were memory-bound, a smaller model often runs significantly faster and cheaper on the same hardware.
It doesn’t require an ASIC, but GPUs/AI accelerators will typically deliver the highest throughput for transformer inference once the model fits comfortably in memory.
- What does the compression rely on?
CompactifAI relies on redundancy in trained transformer weight matrices: large models are often overparameterised, so the same behaviors can be represented with fewer effective parameters.
Instead of generic “zip-style” compression, it uses a model-aware factorization (quantum-inspired tensor networks) to rewrite large matrices into a structured, smaller form while mitigating the accuracy trade-off.
- What prevents others from copying your techniques/process? Analogous to the various compression techniques that are available (e.g. zip, rar, 7z etc)
Multiverse Computing’s proprietary CompactifAI technology is a unique approach to AI model compression, based on research into quantum-inspired tensor networks by cofounder and CEO Roman Orus, and the company’s own research team.
What prevents copycat techniques is the technical know-how required to achieve such high rates of compression without sacrificing accuracy.
CompactifAI can reduce model size by up to 95% with only a 2-3% accuracy loss, compared to the industry standard of 20-30% accuracy loss after just 50-60% compression.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
https://cdn.mos.cms.futurecdn.net/KCJxM3ey3eCaGs2Kdr7kcL-2223-80.jpg
Source link
desire.athow@futurenet.com (Desire Athow)
