Llms do not require powerful servers: Investigators from MIT, KAAST, ISTA, and yandex introduces a new AI in large languages ​​in large numbers of languages ​​without greater degree of high quality

• HIGGS – A new way of oppression of large models of languages ​​developed in conjunction with the Yandex Research teams, Mit, Kaaust and ITTA.
• Higgs make it possible to stress llms without additional data or promometer.
• Unlike other ways of oppression, higgs do not need a special hardware and strong gpu. Models that can be separated directly from the smartphone or laptop in a few minutes without the loss of the important quality.
• The manner has already been used to measure the popularity of lllama 3.1 and family models, and deep family models and QWEN.

Yandex Research team, along with the researchers from the Massachusetts Institute of Technology (Mit), the Austrian Institute of Science Nking Abdullah of Science and Technology (Kaunt), improving the huge models of languages ​​without much quality.

Eve earlier, sending large models of language on mobile devices or laptops include the minimum range of size – it takes anywhere from hours to the weeks – to keep industrial servers – to maintain good quality. Now, the construction of energy can be eliminated in the minutes in a minute directly in the smartphone or laptop without a field of industry or powerful GPU.

Higgs lowered the barrier to the checkpoint and sending new models on consumer distance devices, such as homes and smartphones by removing the need for industrial electricity power.

The new variable approach is the commitment of the company in making large languages ​​of languages ​​available to everyone, from largest players, SMBs, and non-profit organizations with certain donors, developers and investigators. Last year, Yandex investigators work together with large scientific universities and technology to introduce two alternative LMM Coopry systems: To add the number of large language models (AQLM) and PV. Included, these methods can reduce the size of the model in 8 times while storing 95% answer quality.

Delivery down the idlist of the acquisition of llm

Large types of languages ​​require the resources of the integration of the combination, which make themselves out of reach and expensive. This is also the case with open models, such as Deepseek R1 famous R1, which cannot be easily distributed even to the most developed servers designed for example training and other machine learning activities.

As a result, access to these powerful models accordingly are limited to several selected organizations and infrastructure required for the billing and charge, despite their public access.

However, Higgs can open a wide range of widespread. Developers now reduce the model size without giving up quality and running them on the most expensive devices. For example, this approach can be used to press the Deepseek R1 with 671 parameters with the 400 parameters, before they lose (pressed) for a great loss in important quality. This amount of prices opens new ways to use llms throughout all the various fields, especially in the oppressed resources. Now, the Startups and independent developers can gain the strength of pressed models to build new products and services, while cutting costs from expensive machines.

Yandex has already used Higgs in Protetype and speed up product development, as well as the assessment of the concept, as pressed models enable immediate examine.

About the way

Hitzing Hivgs (Havamard Havamard with Gaussian Ms-Ortic Grids) Press large models of language without requiring additional data or decorations are easily available and efficient for various apps and devices. This is especially important when there is a lack of proper balance of model. The method provides a balance between the quality of model, size, and difficulty reduction, which makes it possible to use models on various devices such as smartphones and consumer laptops.

Higgs tested in Llama Models 3.1 and 3.2-Family, and QWen family models. Tests indicate that Higgs filter for other data incoming amounts, including NF4 (4-bit chain) and HQQ (half-quadratic quadization), according to high quality measure.

The engineers and investigators have access to the method in the line of face or testing the research paper, which is available from the arxiv. At the end of this month, the party will introduce their paper in NAACL, one of the highest conferences in the world in Ai.

Continual commitment to improving science and doing well

This is one of the papers of yander wandex presented in the thick model. For example, the Group and PV group also create a service that allows users to use 8B model on a normal PC or smartphone in the displayed display, or without high computer power.

In addition to the LLM camp, the yandeX is detected for a few tools that are expanding resources used in the LLM training. For example, the YAFSDP library is speeding up the llm training as 25% and reduce the 20% training gPU services.

Earlier this year, open yandex developers, the continuous test tool for the actual time and servers analysis and apps. The Performator highlights the non-functional code and provides effective understanding, which helps companies reduce the cost of up to 20% infrastructure. This can translate the potential sector or even billions of dollars a year, depending on the size of the company.

Survey The paper. All credit for this study goes to research for this project. Also, feel free to follow it Sane and don't forget to join ours 85k + ml subreddit. Note: Because of the yandex team of the leadership of the thinking / resources of this topic. Yandex group supported financially for this content / article.