Best open source LLMs: Vicuna-13B, BLOOM, MPT-7B (2024)

Here we bring the best open source Large Language Models (LLMs) that will shape AI development in the coming years. You can use them for free or for a small fee.

The LLM market is constantly growing and specializing, and while large AI models such as ChatGPT or Google Gemini offer increasingly mind-blowing capabilities to a wide range of users, there are more and more narrowly focused projects that may not have such broad capabilities, but can be much more effective when solving specific problems. And importantly, many of these alternative LLMs can even be used for commercial purposes. Modern solutions such as LM Studio make it much easier to use free AI models, and to work with them you only need an ordinary Mac M1/M2/M3 or PC with AVX2 support. All this makes open models of large languages accessible not only to large companies, but also to small businesses, startups and entrepreneurs.

Here are the most notable open source LLMs:

Vicuna-13B

We put Vicuna-13B in first place because it is perhaps the most spectacular and daring project among open source LLMs. This outperforms LLaMA, Stanford Alpaca and many other models more than 90% of the time and almost reaches the capabilities of OpenAI ChatGPT and Google Bard. Meanwhile, no more than $300 was spent on his training process, which perfectly demonstrates the accessibility and openness of LLM technology.

Vicuna-13B was built from user conversations obtained from ShareGPT and ChatGPT and fine-tuned using LLaMA, resulting in high response quality. This is an open source model, and any user can test its capabilities on its demo page. The code and weight, as well as online demonstration are publicly available for non-profit use. Vicuna-13b training will cost the user $ 300.

Llama 2

Llama 2 was a huge step forward from Llama 1, Meta's first AI model. Thanks to a partnership with Microsoft, Llama 2 received integration into Azure and Windows, bringing it closer to millions of users and developers. In addition, it is supported by mobile devices based on Qualcomm Snapdragon.

This model has 70 billion parameters and offers fine-tuning capabilities across multiple platforms. Llama 2 is available in two main versions - the main one and the Llama 2-Chat bot created for two-way conversations. These models provide up to 70 billion parameters.

BLOOM

BLOOM is a powerful large language model (LLM) that serves autoregressive text generation capable of expanding a given text prompt. In simple terms, this language model is designed to create complete texts from given prompts. Built on 176 billion parameters, BLOOM is capable of generating coherent text in 46 languages and 13 programming languages. For almost all of them BLOOM is the first LLM with over 100B parameters ever created.

BLOOM was developed as part of the Hugging Face platform project, which involved activist developers from over 70 countries, and is one of the most transparent and accessible LLMs, providing absolutely free access to source code and training data. Such transparency allows us to constantly develop and perfect this language model.

MPT-7B

MPT-7B (MosaicML Pretrained Transformer) is a super fast GPT-style transformer model developed by MosaicML Foundations and trained on 1 trillion text and code tokens. It also offers open source code and, in addition, the possibility of commercial licensing, making it a valuable tool for companies and public institutions.

MPT-7B can be used to create analytics, forecasts and make strategic decisions. In addition to the main model, there are also variants with narrower applications. MPT-7B-Chat is designed for creating dialogues, MPT-7B-StoryWriter-65k+ allows you to create long analytical chains, and MPT-7B-Instruct is designed for short instructions.

Falcon

Like GPT, the Falcon model is also built on the principle of an autoregressive decoder, that is, it predicts the next token based on previous tokens. The advantage of this model is that it uses very high-quality and carefully filtered content from the Internet. It is trained on an impressive amount of internet data of one trillion tokens and has 40 billion parameters, making it superior to GPT-3.

The main model, Falcon-40B, as well as special variants Falcon-7B, Falcon-7B-Instruct and Falcon-7B-Instruct are open source and available to all users. Falcon 40B LLM works with English, German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian, Czech, and Swedish languages.