GenAI Model Updates: Latest Developments from Meta, Mistral, Apple, Google, OpenAI, and more

Meta's Llama 3.1

Not entirely open-source, but close enough. How open it is?

One of the key improvements in Llama 3.1 is its massive training dataset. It boasts a massive training dataset of 15 trillion tokens, a huge leap from the previous 1.5 trillion. This broader range of text data makes it a more knowledgeable and capable model.

The architecture and training objectives have also been enhanced, allowing the model to better understand nuanced language and generate more coherent text.

Plus, it achieves all this with fewer parameters - 70 billion, down from 100 billion - making it a more efficient model overall. Read more

Mistral Large 2

Mistral AI, another company competing with Meta to develop the best open-source models, has released Mistral Large 2.

According to their website, this model boasts 123 billion parameters, making it quite substantial. Their shared documentation shows that Mistral Large 2 outperforms Llama 3.1 70B in math performance.

You can find more benchmarks on their website. Overall, Mistral Large 2 performs on par with the best state-of-the-art closed and open-source models across various benchmarks. Read more

Apple's AI Models

They are a family of AI models, including a 7 billion parameter model and a 1.4 billion parameter model, as part of the DataComp for Language Models project.

Open-source or not: Yep! Apple has released the model weights, training code, and pretraining dataset, making the project truly open-source.

High-Performing: The 7B model has outperformed Mistral-7B and is closing in on other leading open models, including Llama 3 and Gemma.

Commercial Use: The smaller 1.4B model has been released under Apache 2.0, allowing for commercial use, distribution, and modification. Read more

??Bonus: Two free webinars are coming in less than 48 hours!

??? How to make full-stack MongoDB and JSON apps 100X faster

??? Building Fast GenAI Apps on Mongo Data with SingleStore Kai

Google Gemini's Upgrade: Enhanced Conversational AI Capabilities

This represents a substantial improvement, offering enhanced capabilities, faster response times, and expanded access. Its focus on contextual understanding, attention mechanisms, and reinforcement learning makes it a robust and adaptable conversational AI solution. Read more

Not entirely open-source. While Google has released some components and tools related to Gemini, the full model architecture and weights are not publicly available.

SearchGPT Prototype: AI-Powered Search Engine

This is an experimental AI-powered search engine developed by OpenAI, leveraging the capabilities of the GPT model. Some key features and aspects:

• AI-Powered Search: It is specifically designed for search, allowing users to interact with it like a search engine. ChatGPT with internet access can also search the internet, but it's not its primary function.

• Conversational Search: It is optimized for conversational search, allowing users to ask follow-up questions and engage in a dialogue.
• Indexing and Retrieval: SearchGPT Prototype has its own indexing and retrieval system, enabling it to quickly find and generate responses from a massive dataset Read more

Kling AI Video Generator Now Available Globally

Kling AI (proprietary model), developed by Kuaishou, has launched globally, offering a multilingual AI video generator platform that supports text-to-video and image-plus-text-to-video generation. Key features include:

• Multilingual support: High-quality results for prompts in English and Mandarin
• Video resolutions: 960x960 square format, 720p landscape, and 720p portrait mode
• Camera control: Static, horizontal, vertical, pan, tilt, or rolling camera movements
• Fine-tuning: Balance between creativity and relevance, with negative prompts for accuracy
• Content restrictions: No violent scenes, famous people, or NSFW content Read more

Sakana AI's Image Models

Sakana AI, a Tokyo-based startup, has released two image generation models, Evo-Ukiyoe and Evo-Nishikie, designed to generate traditional Japanese ukiyo-e artwork.

These models use AI to create images from text and image prompts, focusing on Japan's historic art form that flourished between the 17th and 19th centuries. Read more

Open-source or not: Yep, they are open-source. But please note 1) they are not intended for commercial use or deployment in mission-critical environments. 2) The models are considered experimental prototypes, and their performance and outcomes are not guaranteed.

This issue is brought to you in partnership with Auxi.

auxi is an AI-driven solution that streamlines presentation slide creation, offering expert assistance in content suggestions, design, and data analysis.

Trusted by over 350 institutions, including top global consulting and investment banking firms. They currently offer free trials, start here.