Meta New Models - SAM 2.1, Spirit LM, MEXMA and More
Summary
Meta has recently introduced the Meta Spirit LM, an open-source language model capable of accepting and generating text and speech inputs. This model, utilizing three additional speech tokenizers, showcases improved speech quality and expressiveness similar to GPT-4. Meta's ongoing research projects, such as Postquantum and Metal Lingua, aim to enhance security, boost computational efficiency, and improve performance of their models.
Meta's New Models Release
Meta has released a variety of different models focused on various applications, including new research models and datasets from Meta's fundamental research lab Meta AI Research (Fair) aiming at achieving advanced machine intelligence.
Meta Spirit LM Model
Meta has introduced the new model called Meta Spirit LM, an open-source language model that can accept and generate both text and speech. The model is equipped to accept speech input and can convert speech into text. It demonstrates similar speech expressiveness to GPT-4 and ensures better speech quality compared to the previous 7 billion model.
Technical Details of Meta Spirit LM
The Meta Spirit LM model consists of three additional speech tokenizers for processing pitch and style of input speech, and it uses the Lama 27 billion model. Users need to apply for access to the model's weights, which offer improvements in inference speed and overall performance.
Research Projects by Meta
Meta is engaged in multiple research projects, including Postquantum for enhancing security against AI, Metal Lingua for lightweight language model training, and Materials 2024 for inorganic data and model research. These projects aim to improve performance and boost computational efficiency.
Crosslingual Sentence Encoder
Meta's research on a pre-trained crosslingual sentence encoder using token-level objectives shows promising results, particularly in crosslingual translations. The encoding enables better translations across various languages, enhancing the model's effectiveness.
Generative Reward Models
Meta has developed strong generative reward models and self-taught synthetic preference data to train annotations, creating reasoning traces for model evaluations. The models are accessible for usage but require permissions for access, offering improved performance and outperforming larger models like 405b Instruct and Gemini Pro.
FAQ
Q: What is the new model introduced by Meta called?
A: The new model introduced by Meta is called Meta Spirit LM.
Q: What capabilities does the Meta Spirit LM model have?
A: The Meta Spirit LM model can accept and generate both text and speech. It can convert speech into text and has three additional speech tokenizers for processing pitch and style of input speech.
Q: What is the Meta Spirit LM model equipped with for speech processing?
A: The Meta Spirit LM model is equipped with the Lama 27 billion model and three additional speech tokenizers for processing pitch and style of input speech.
Q: What research projects is Meta engaged in?
A: Meta is engaged in multiple research projects, including Postquantum for enhancing security against AI, Metal Lingua for lightweight language model training, and Materials 2024 for inorganic data and model research.
Q: What are the objectives of Meta's research on a pre-trained crosslingual sentence encoder?
A: The research aims to achieve better translations across various languages and enhance the model's effectiveness through token-level objectives.
Get your own AI Agent Today
Thousands of businesses worldwide are using Chaindesk Generative
AI platform.
Don't get left behind - start building your
own custom AI chatbot now!