Today's Open Source (2024-08-05): infly-ai Releases 34B Pre-trained Model with Bilingual Support and 32K Context
Discover top AI open-source projects: INF-34B, Grounded SAM 2, Llama Coder, LangGraph Studio, Autogen_GraphRAG_Ollama, and LLMDocParser. Explore cutting-edge tech now!
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: INF-34B
INF-34B is a model with 34 billion parameters and 32K context, trained on about 3.5T bilingual Chinese-English corpus.
Compared to other open-source models of similar size, INF-34B performs excellently on OpenCompass evaluations and shows great potential in finance and healthcare.
Moreover, the quantized INF-34B can run on a 24GB VRAM GPU with almost no accuracy loss, making it suitable for commercial applications, especially in low-resource scenarios.
https://huggingface.co/infly/INF-34B-Chat
https://huggingface.co/infly/INF-34B-Base
Project: Grounded SAM 2
Grounded SAM 2 is an open-source project by IDEA combining Grounding DINO and SAM 2, aimed at solving complex visual tasks.
The project simplifies code implementation, making deployment easier for developers.
Grounded SAM 2 supports object tracking in images and videos, offering various pre-trained models and demos.
https://github.com/IDEA-Research/Grounded-SAM-2
Project: Llama Coder
Llama Coder is an open-source project using Llama 3.1 to build Claude-like Artifacts.
The project can generate small applications from a prompt, with a tech stack including Next.js, Tailwind, and Sandpack.
https://github.com/Nutlope/llamacoder
Project: LangGraph Studio
LangGraph Studio is a desktop app for local development and debugging of LangGraph applications.
It offers visual charts and state editing to help users understand agent workflows and speed up iteration.
LangGraph Studio integrates with LangSmith and is free for all LangSmith users during the Beta phase.
https://github.com/langchain-ai/langgraph-studio
Project: Autogen_GraphRAG_Ollama
Autogen_GraphRAG_Ollama integrates GraphRAG, AutoGen, Ollama, and Chainlit into a local multi-agent RAG bot.
The project uses local LLMs like Ollama for free and offline embedding and inference, providing an interactive UI for handling continuous conversations, multi-threading, and user input settings.
https://github.com/karthik-codex/Autogen_GraphRAG_Ollama
Project: LLMDocParser
LLMDocParser is a package for parsing PDFs and using LLMs to analyze their content.
Improving on the gptpdf concept, it can identify text and non-text areas in PDFs and input the results into a multimodal model for analysis.
Its main features include layout analysis, region type recognition, and text block extraction, suitable for tasks requiring high-precision PDF parsing and content analysis.