laitimes

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

author:New Zhiyuan

Editor: Yong Yong Qiao Yang

On May 14th, the open-source large language model Falcon 2 was released, with performance surpassing Llama 3, and the news was on the first place on the Hacker News hot list. With the return of the "Falcon", will the open source universe usher in a new overlord?

In the past few days, the eyes of the world seem to be attracted by the GPT-4o released by OpenAI, and at the same time, OpenAI's challengers are also making history at the same time.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

On May 14, the Technology Innovation Institute (TII), part of the Advanced Technology Research Council (ATRC) in Abu Dhabi, unveiled a new generation of the Falcon 2 model.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

Falcon 2 11B is open for access, and the Falcon 2 11B VLM will open at 12 noon on May 14

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

The new generation of "Falcon" (Falcon means Falcon) returned to the arena and quickly reached the top of the HN Hot List.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

Last year, the Falcon debuted with a stunning performance, surpassing the Llama by a crushing advantage.

According to the Hugging Face rankings, the Falcon 2 11B once again outperforms the Llama 3 8B and is on par with the leading Gemma 7B, with an average score of just 0.01.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

Falcon 2包括两个版本:Falcon 2 11B和Falcon 2 11B VLM。

The former is a large language model trained on 5.5 trillion tokens, while the latter is a "vision-to-language" model developed on top of it.

As an open-source large language model, Falcon 2 not only has multilingual capabilities, but also supports multimodality, which can seamlessly convert visual input into text output, which is the first of its kind in today's large model market.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

All you need to do is upload an image to the user interface and enter a prompt to get a text description of the image

Multilingual + multimodal

The Falcon 2 11B is multilingual and can quickly respond to assignments in English, French, Spanish, German, Portuguese, and a variety of other languages.

As a visual language model, the Falcon 2 11B VLM is capable of recognizing and interpreting various visual information from the environment, so it has a wide range of application scenarios in fields such as healthcare, finance, e-commerce, education, and law.

For example, it is used for tasks such as document management and contextual indexing, and its "sight-to-language" function is expected to help the visually impaired in the future.

In addition, the Falcon 2 runs efficiently with only one GPU and can be easily deployed on laptops.

In order to accelerate the commercialization of Falcon, in November last year, ATRC established a new company, AI71

Based on the Falcon model, AI71 has already entered the United Arab Emirates in education, healthcare, law and other fields, and plans to expand beyond the UAE soon. This upgrade of Falcon provides strong support for its commercialization process.

Stick to the open-source Falcon

TII's AI models have been among the best in the world, and the new Falcon 2 11B model has reaffirmed TII's position.

What's more, both model versions of this update have been open-sourced and can be used by developers around the world without restrictions to develop and deploy on top of them. The excellent performance of the model means that TII is expected to compete with Meta for the hegemony of the open source universe.

Faisal Al Banni, Secretary-General of the Abu Dhabi Advanced Technology Research Council (ATRC), pledged to continue to drive open source innovation to help developers around the world realize their AI journeys.

Falcon is back after a year! 11 billion parameters and 5.5 trillion tokens, surpassing Llama 3 in performance

Dr. Hakim Hacid, Principal Investigator at TII, also said, "Developers are realizing the myriad benefits of smaller, more efficient models······ This is the next emerging megatrend. In addition, the visual-to-language capabilities of Falcon 2 open up new prospects for the ease of use of AI, providing users with a transformative way to interact from image to text."

TII also revealed that it is working on the next generation of Falcon 2 models, which will be launched shortly, and the performance of these models will be further enhanced through expert hybrid (MoE) technology.

The battle between open source and closed source is undecided. Regardless, TII has become a powerhouse in the AGI space, and behind the top-down synergy of the UAE, which is pushing for the development of artificial intelligence, which is seen as the next economic pillar after oil.

Falcon doesn't just exist as a "serious participant", but an ambitious "falcon" that hovers around with wings.