Open Source Today (2024-09-10): DeepSeek-Coder-V2 Released, Now Supports 338 Languages and 128K Context
Explore top AI open-source projects: DeepSeek-Coder-V2, CodeFuse-muAgent, Anything-LLM, Flux Gym, KubeAI, and Tomato. Boost your coding and AI solutions today!
Here are some interesting AI open-source models and frameworks I wanted to share today:
Project: DeepSeek-Coder-V2-Instruct-0724
DeepSeek released their DeepSeek-Coder-V2-Instruct-0724, which ranks 2nd on the Aider LLM leaderboard, surpassing DeepSeek V2.5.
DeepSeek-Coder-V2 is an open-source Mixture-of-Experts (MoE) code model. It performs on par with GPT4-Turbo on code-specific tasks.
This model is further pre-trained from DeepSeek-V2’s intermediate checkpoint with 6 trillion more tokens. It boosts coding and math reasoning skills while maintaining solid performance in general language tasks.
Supported programming languages expanded from 86 to 338, and the context length increased from 16K to 128K.
https://huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct-0724
Project: CodeFuse-muAgent
CodeFuse-muAgent is an innovative agent framework powered by a knowledge graph engine. It combines multi-agent systems, function calls, and code interpreters.
With drag-and-drop features and simple text, large language models can carry out complex procedures under human guidance.
This framework has been proven in various complex DevOps scenarios at Ant Group.
https://github.com/codefuse-ai/CodeFuse-muAgent
Project: Anything-LLM
AnythingLLM is a full-stack app that lets users turn any document, resource, or content into context for any large language model (LLM).
It supports multiple LLMs or vector databases, multi-user management, and permission settings.
Users can run it locally or host it remotely to have smart conversations with any provided document.
https://github.com/Mintplex-Labs/anything-llm
Project: Flux Gym
Flux Gym is a simple Web UI for training FLUX LoRA in low-memory environments (12GB/16GB/20GB).
It combines AI-Toolkit's Gradio UI for the frontend with Kohya Scripts on the backend, offering a flexible and easy-to-use interface for LoRA training.
https://github.com/cocktailpeanut/fluxgym
Project: KubeAI
KubeAI is a private OpenAI alternative running on Kubernetes. It supports various open-source models (like LLMs, Whisper) and offers an OpenAI-compatible API.
KubeAI supports multi-platform setups (CPU, GPU) and scales automatically based on the workload. It also includes a chat interface (OpenWebUI) and message integration (Kafka, PubSub).
https://github.com/substratusai/kubeai
Project: Tomato
Tomato is a proof-of-concept steganography tool that uses Minimum Entropy Coupling (MEC) to hide encrypted messages in natural language.
By aligning the probability distribution of hidden messages with the LLM-generated cover text, it makes the embedded messages statistically indistinguishable from normal language, making them hard to detect.
