Over the past month, there have been some very interesting and significant events in the world of Large Language Models (LLM).
Major companies have released fresh versions of their models. First, Google launched two new models, Gemini: Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002.
Key Features:
- More than a 50% price reduction for the 1.5 Pro version
- Results are delivered twice as fast with three times lower latency
The main focus has been on improving performance and speed and reducing costs for models intended for industrial-grade systems.
Details here
DataGemma
Google also released the DataGemma model.
It includes a series of finely-tuned Gemma 2 models that assist language models (LLM) in accessing numerical and statistical data.
In particular, they offer a new approach called Retrieval Interleaved Generation (RIG), which reliably integrates public statistical data from Data Commons into LLM responses.
RIG is a tool-based approach that can interleave statistical tokens with natural language questions suitable for retrieval from Data Commons.
To achieve this capability, they fine-tune LLM on a dataset of instruction and answer pairs generated using Gemini 1.5. The RIG approach improves factuality from 5-7% to approximately 58%.
Llama 3.2
Meta is not falling behind and has released Llama 3.2: a lightweight model for devices, computer vision (Llama Vision), and much more! Meta released 10 new models ranging from 1B (text only) to 90B multimodal (text + image).
Brief Overview:
- Llama 3.2 Vision: multimodal (text + image to text) models sized 11B and 90B based on the text models Llama 3.1, trained on 6 billion text-image pairs.
- Llama 3.2 Edge: multilingual text models at 1B and 3B for efficient local deployment.
- All Llama 3.2 models support a context length of 128k tokens.
- Knowledge distillation and pruning were used from 8B/70B to train the 1B/3B models.
- Llama Guard 3.2: 2 new, improved guard models with multimodal support.
- The Llama 3.2 3B model is comparable to the Llama 3.1 8B in IFEval, indicating strong use cases for local RAG applications or agents.
- Multimodal models are restricted for individuals and companies within the European Union.
- Available and integrated into the Hugging Face ecosystem from Hub to Transformers and TGI.
More details in the official announcement: https://go.fb.me/8ar7oz
Download Llama 3.2 models: https://go.fb.me/7eiq2z
Llama-3.2’s capabilities in computer vision are expected to surpass those of GPT-4o-mini.
Multimodal capabilities are becoming essential as we move towards the next phase of more autonomous systems.
The Llama 3.2 models at 1B and 3B, which support context lengths of up to 128K tokens, also appear impressive and are useful for applications on edge devices and mobile platforms.
The Llama Stack API seems to have received several enhancements, simplifying the development of agent applications. Now, there is a CLI for Llama, client code in languages like Python and Node, Docker containers, and several distribution options.
OpenAI Academy
Meanwhile, OpenAI has delighted us with the academy‘s opening. The academy aims to foster innovation by investing in developers and organizations using AI, starting with low—and middle-income countries.
The OpenAI Academy program will provide:
- Training and technical support: Assistance from OpenAI experts for developers and targeted organizations using AI.
- API credits: Distribution of an initial $1 million in API credits to expand access to OpenAI models, allowing participants to create and deploy innovative applications.
- Community building: Formation of a global network of developers for collaboration, knowledge sharing, and collective innovation.
- Competitions and incubators: Partnerships with philanthropists to invest in organizations addressing critical issues in their communities.
Currently, there are no details on participation.