Today's Open Source (2024-07-25): Mistral AI Releases Large 2 with 128k Context
Explore the latest AI open-source models: Mistral Large 2 with 123B parameters, AgentScale for AI agent orchestration, Meta's Llama toolkit, and more.
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: Mistral Large 2
Mistral AI has released its latest model, Mistral Large 2, with 123B parameters, designed for single-node inference, achieving high throughput on a single node with a 128k context window.
It excels in code and math reasoning, supporting dozens of languages including Chinese, Japanese, and Korean, and over 80 programming languages like Python, Java, C, C++, JavaScript, and Bash.
It also supports function calls and structured output natively. Its pre-trained version reached 84.0% accuracy on MMLU.
https://huggingface.co/mistralai/Mistral-Large-Instruct-2407
https://mistral.ai/news/mistral-large-2407/
Project: AgentScale
AgentScale is a microservices-based next-gen agent orchestration framework.
It redefines building and deploying AI agents with smart routing, stateful conversation management, and scalable architecture.
AgentScale supports personalized, context-aware interactions, offers a unified API gateway, and simplifies the integration and management of new agent services.
https://github.com/M1n9X/AgentScale
Project: Llama Models
Meta has open-sourced a toolkit for Llama models, providing easy access to the entire Llama1~3 series, helping developers, researchers, and businesses build AI applications.
https://github.com/meta-llama/llama-models
Project: lmms-finetune
lmms-finetune is a minimal codebase for fine-tuning large multimodal models, supporting models like LLaVA-1.5, Qwen-VL, LLaVA-Interleave, LLaVA-Next-Video, and Phi-3-Vision.
It aims to provide a unified, concise structure, making fine-tuning these models simpler and more efficient.
https://github.com/zjysteven/lmms-finetune
Project: Llama Agentic System
The Llama Agentic System is an open-source system from Meta for performing "agentic" tasks, built on the Llama 3.1 model.
It can decompose tasks and execute multi-step reasoning, with built-in tool usage like search and code interpreters.
The system can also call unseen tool definitions through zero-shot learning.
The project shifts safety evaluations from model-level to system-level to ensure broad controllability and adaptability in various use cases with different safety requirements.
https://github.com/meta-llama/llama-agentic-system
Project: MINT-1T
MINT-1T is an open-source multimodal dataset with one trillion text tokens and 3.4 billion images, about ten times the size of existing open-source datasets.
It also includes previously unused resources like PDFs and ArXiv papers.