Today's Open Source (2024-11-01): Meta Releases MobileLLM Model Code
Explore new AI open-source projects: MobileLLM, SmolLM2, Genmoai-smol, SandboxFusion, ComfyUI EcomID, and TrustGraph Engine. Innovative models and tools today!
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: MobileLLM
MobileLLM is a billion-parameter language model launched by Meta that is optimized for on-device usage scenarios.
This project integrates various design elements, such as the SwiGLU activation function, deep-thin architecture, embedding sharing, and grouped-query attention mechanism, to achieve a high-quality language model with fewer than one billion parameters.
MobileLLM-125M/350M has shown significant accuracy improvements in zero-shot commonsense reasoning tasks compared to previous best models of similar scale, and it greatly outperforms small LLMs with the same parameter count in chat and API call tasks.
https://github.com/facebookresearch/MobileLLM
Project: SmolLM2
SmolLM2 is a compact language model available in 1.7B, 360M, and 135M parameter sizes.
The model demonstrates significant progress in instruction following, knowledge reasoning, and mathematical capabilities.
Through supervised fine-tuning and reinforcement learning from human feedback, this model supports tasks such as text rewriting, summary generation, and function calling.
https://huggingface.co/HuggingFaceTB/SmolLM2-135M
https://huggingface.co/HuggingFaceTB/SmolLM2-360M
https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B
Project: Genmoai-smol
Genmoai-smol is an open-source video generation model optimized to run on a single GPU node, suitable for environments with limited VRAM.
The project aims to provide high-fidelity motion and strong prompt adherence, bridging the gap between open and closed video generation systems.
Users can generate videos directly via Gradio UI or command-line interface.
https://github.com/victorchall/genmoai-smol
Project: SandboxFusion
SandboxFusion is a secure sandbox for running and evaluating code generated by large language models (LLMs).
The project supports multiple programming languages and provides an online evaluation feature, suitable for code execution needed in evaluation and reinforcement learning datasets.
The project aims to provide developers with a safe, reliable environment to test and validate code.
https://github.com/bytedance/SandboxFusion
Project: ComfyUI EcomID
ComfyUI EcomID is a project for enhancing portrait rendering, aiming to provide a more realistic and aesthetically pleasing appearance while ensuring semantic consistency and higher internal ID similarity.
The project natively supports SDXL-EcomID and is fully integrated with ComfyUI, primarily applied in facial recognition and image processing fields.
https://github.com/alimama-creative/SDXL_EcomID_ComfyUI
Project: TrustGraph Engine
TrustGraph Engine is a platform for developing knowledge-driven intelligent agents, providing all the tools, services, graph storage, and vector databases needed to deploy reliable, scalable, and accurate AI agents.
Its features include batch document ingestion, automated knowledge graph construction, automated vectorization, model-agnostic LLM integration, and Retrieval-Augmented Generation (RAG) that combines knowledge graphs and vector databases.
The platform supports enterprise-level reliability, scalability, and modularity, ensuring data privacy through Ollama and Llamafile.
