Large Language Models No Longer Require Powerful Servers
Scientists from Yandex, HSE University, MIT, KAUST, and ISTA have made a breakthrough in optimising LLMs. Yandex Research, in collaboration with leading science and technology universities, has developed a method for rapidly compressing large language models (LLMs) without compromising quality. Now, a smartphone or laptop is enough to work with LLMs—there's no need for expensive servers or high-powered GPUs.
This method enables faster testing and more efficient implementation of new neural network-based solutions, reducing both development time and costs. As a result, LLMs are more accessible not only to large corporations, but also to smaller companies, non-profit laboratories and institutes, as well as individual developers and researchers.
Previously, running a language model on a smartphone or laptop required quantising on an expensive server—a process that could take anywhere from a few hours to several weeks. Quantisation can now be performed directly on a smartphone or laptop in just a few minutes.
Challenges in implementing LLMs
The main obstacle to using LLMs is that they require considerable computational power. This applies to open-source models as well. For example, the popular DeepSeek-R1 is too large to run even on high-end servers built for AI and machine learning workloads, meaning that very few companies can effectively use LLMs, even if the model itself is publicly available.
The new method reduces the model's size while maintaining its quality, making it possible to run on more accessible devices. This method allows even larger models, such as DeepSeek-R1 with 671 billion parameters and Llama 4 Maverick with 400 billion parameters, to be compressed, which until now could only be quantised using basic methods and resulted in significant quality loss.
The new quantisation method opens up more opportunities to use LLMs across various fields, particularly in resource-limited sectors such as education and the social sphere. Startups and independent developers can now implement compressed models to create innovative products and services without the need for costly hardware investments. Yandex is already applying the new method for prototyping—creating working versions of products and quickly validating ideas. Testing compressed models takes less time than testing the original versions.
Key details of the new method
The new quantisation method is named HIGGS (Hadamard Incoherence with Gaussian MSE-Optimal GridS). It enables the compression of neural networks without the need for additional data or computationally intensive parameter optimisation. This is especially useful in situations where there is not enough relevant data available to train the model. HIGGS strikes a balance between the quality, size, and complexity of the quantised models, making them suitable for use on a variety of devices.
The method has already been validated on the widely used Llama 3 and Qwen2.5 models. Experiments have shown that HIGGS outperforms all existing data-free quantisation methods, including NF4 (4-bit NormalFloat) and HQQ (Half-Quadratic Quantisation), in terms of both quality and model size.
Scientists from HSE University, the Massachusetts Institute of Technology (MIT), the Austrian Institute of Science and Technology (ISTA), and King Abdullah University of Science and Technology (KAUST, Saudi Arabia), all contributed to the development of the method.
The HIGGS method is already accessible to developers and researchers on Hugging Face and GitHub, with a research paper available on arXiv.
Response from the academic community, and other methods
The paper describing the new method has been accepted for presentation at one of the largest AI conferences in the world—the North American Chapter of the Association for Computational Linguistics (NAACL). The conference will be held from April 29 to May 4, 2025, in Albuquerque, New Mexico, USA, and Yandex will be among the attendees, along with other companies and universities such as Google, Microsoft Research, and Harvard University. The paper has been cited by Red Hat AI, an American software company, as well as Peking University, Hong Kong University of Science and Technology, Fudan University, and others.
Previously, scientists from Yandex presented 12 studies focused on LLM quantisation. The company aims to make the application of LLMs more efficient, less energy-consuming, and accessible to all developers and researchers. For example, the Yandex Research team has previously developed methods for compressing LLMs, which reduce computational costs by nearly eight times, while not significantly compromising the quality of the neural network’s responses. The team has also developed a solution that allows running a model with 8 billion parameters on a regular computer or smartphone through a browser interface, even without major computational power.
See also:
Neural Network Maps as a Method for Constructing Mathematical Models
Scientists from HSE University–Nizhny Novgorod and the Institute of Physics Belgrade, Serbia, are jointly exploring the application of machine learning techniques and neural networks to the study of nonlinear dynamics. Natalya Stankevich, Leading Research Fellow at the Laboratory of Topological Methods in Dynamics of the Faculty of Informatics, Mathematics, and Computer Science at HSE University–Nizhny Novgorod, spoke to the HSE News Service about this international project.
HSE Scientists Develop Method to Compress Large Language Models Without Losing Quality
Researchers from the AI and Digital Science Institute at the HSE Faculty of Computer Science have developed a new compression method for large language models such as GPT and LLaMA that reduces their size by 25–36% without additional training or significant loss of accuracy. This is the first approach to use mathematical transformations—specifically, rotations of model weights—to make models more amenable to compression with structured matrices. The study results have been published in ACL Findings 2025. The code is available on GitHub.
Machine Learning Models Can Help Reduce Volatility and Boost Stock Market Returns
The use of machine learning models makes it possible to achieve greater accuracy in predicting risks in the Russian stock market compared to classical econometric approaches. The predictive power of these models increases by 23%, while the average investor’s return can reach up to 13% per annum. These conclusions were drawn by Nikita Lysenok from the Department of Financial Market Infrastructure at the HSE Faculty of Economic Sciences. The paper has been published in Fundamental and Applied Mathematics.
Pocket Money, Personal Interest, and Family Practices: What Shapes Students’ Economic Literacy?
University students' economic literacy depends not only on their field of study but also on their interest in economics, the learning environment, and family financial practices. For example, students who received pocket money irregularly tend to perform better on economic literacy tests than their peers who received financial support on a regular basis. These findings come from a study conducted by HSE University involving more than 1,100 students from five Russian universities. The findings have been published in Cakrawala Pendidikan.
HSE Study Reveals Imbalance in the Generative AI Market
Researchers at HSE University analysed how effectively the global generative artificial intelligence market converts investment into real revenue, concluding that AI is currently developing faster than it is paying off. The results have been published in the journal Foresight and STI Governance.
HSE Scientists Train Neural Network to 'Hear' Faults in Electric Motors
Researchers at the AI and Digital Science Institute of the HSE Faculty of Computer Science have developed a new method—the Signature-Guided Data Augmentation (SGDA) framework—that achieves 99% accuracy in motor fault detection and 86% accuracy in fault classification. The application of this approach can reduce industrial equipment repair costs, minimise downtime, and improve production safety. The study results have been published in Engineering Applications of Artificial Intelligence.
The 'Second Shift' Is Not Why Women Avoid News
Women are more likely than men to avoid political and economic news, but the reasons for this behaviour are linked less to structural inequality or family-related stress than to personal attitudes and the emotional perception of news content. This conclusion was reached by HSE researchers after analysing data from a large-scale survey of more than 10,000 residents across 61 regions of Russia. The study findings have been published in Woman in Russian Society.
Resource Race and Green Transition: Three Unexpected Conclusions from Foresight Centre’s Research on Climate and Poverty
Beneath the surface of green energy—which most people associate with solar panels, electric vehicles, and reduced CO2 emissions—lies a complex web of geopolitical interests, international inequality, and resource constraints. Researchers from the Laboratory for Science and Technology Studies (LST) at the HSE ISSEK Foresight Centre have published a series of articles in leading international journals on hidden and overt conflicts surrounding critically important metals and minerals, as well as related processes in the energy sector.
Immersion in Second Language Environment Influences Bilinguals’ Perception of Emotions
Researchers at the Cognitive Health and Intelligence Centre at the HSE Institute for Cognitive Neuroscience have discovered how bilingual individuals process emotional words in their native (first) and non-native (second) languages. It was found that the link between word meaning and bodily sensations is weaker in a second language than in a first language. However, the more a person is immersed in a language environment, the smaller this difference becomes. The article has been published in Language, Cognition and Neuroscience.
Researchers Find More Effective Approach to Revealing Majorana Zero Modes in Superconductors
An international team of researchers, including physicists from HSE MIEM, has demonstrated that nonmagnetic impurities can help more accurately reveal Majorana zero modes—quantum states considered promising building blocks for quantum computing. The researchers found that these impurities shift the energy levels that typically obscure the Majorana signal, while leaving the mode itself largely unaffected, thereby making its spectral peak more distinct. The study has been published in Research.