Coding & Development

LLMTornado is a comprehensive .NET provider-agnostic SDK designed for developers to build, orchestrate, and deploy AI agents and workflows with ease. It features built-in connectors to over 30 API providers, including Alibaba, Anthropic, Azure, Google, OpenAI, and many more, ensuring broad compatibility without dependencies on first-party SDKs. The library supports first-class local deployments with vLLM, Ollama, or LocalAI, and offers advanced agent orchestration capabilities with concepts like Orchestrator, Runner, and Advancer, including handoffs and parallel execution. LLMTornado accelerates development with its ability to write pipelines once and execute with any provider, and supports fully multimodal inputs and outputs (text, images, videos, documents, URLs, audio). It also integrates cutting-edge protocols like MCP and A2A, and connects to popular vector databases such as Chroma, PgVector, and Pinecone, making it enterprise-ready with guardrails and Open Telemetry support.

LocalAIVoiceChat provides a completely local AI talk experience on your PC, integrating the powerful Zephyr 7B language model with real-time speech-to-text and text-to-speech libraries. It utilizes RealtimeSTT with faster_whisper for transcription and RealtimeTTS with Coqui XTTS for synthesis, allowing for customizable AI personalities and voices. This experimental alpha software requires a GPU with around 8 GB VRAM and specific NVIDIA CUDA or AMD ROCm installations. While not production-ready, it offers a fast and engaging voice-based local chatbot experience, with ongoing updates to improve stability and model performance.

Arcanna.ai delivers trustworthy AI solutions specifically designed for modern Security Operations Centers (SOCs). The platform aims to streamline operations and scale team capacity by automating human judgment in cybersecurity. It features a Decision Layer that triages alerts and enforces decisions, and an Investigation Layer that accelerates investigations using grounded GenAI. Arcanna.ai integrates with existing SIEM, SOAR, NDR, EDR, and XDR workflows, providing explainable decisions, full transparency with logged actions, strict data privacy, and compliance support through recorded decision history and human review points. This approach helps SOC teams achieve faster, more accurate decisions, significantly reducing response times and alert volumes.

MLOpsPython provides a comprehensive set of code samples and pipeline definitions for establishing an MLOps workflow using Azure ML Services and Azure DevOps. This project illustrates how to automate an end-to-end ML/AI workflow, covering continuous integration (CI) and continuous delivery (CD) for machine learning models. It includes tasks for data sanity tests, unit tests, model training on various compute targets, model version management, evaluation, selection, and deployment as a real-time web service. The solution supports staged deployment to QA/production environments and integration testing, making it adaptable for various AI scenarios beyond its scikit-learn diabetes dataset foundation.

DeepResearch, developed by Alibaba-NLP, is an open-source agentic large language model specifically engineered for long-horizon, deep information-seeking tasks. With 30.5 billion total parameters and 3.3 billion activated per token, it demonstrates state-of-the-art performance across various agentic search benchmarks. Key features include a fully automated synthetic data generation pipeline, large-scale continual pre-training on agentic data, and end-to-end reinforcement learning. It supports both ReAct and IterResearch-based 'Heavy' inference paradigms, allowing for rigorous evaluation of intrinsic abilities and maximum performance. The model is available for download and can be integrated via OpenRouter for inference without GPUs.

DeepLearning.scala is an open-source library designed for building complex neural networks using Scala. It supports differentiable programming, allowing users to construct neural networks from mathematical formulas and calculate derivatives for weights. A key differentiator is its ability to create dynamic neural networks, where the structure can change during runtime based on Scala functions and control flows. This enables programmers to build sophisticated networks with familiar coding paradigms. The library also emphasizes functional programming, leveraging Monads and Applicative type classes for composable layers and parallel computations. DeepLearning.scala 2.0 is organized around Dependent Object Type calculus (DOT), providing mixin-able plugins for extending functionality, including algorithms, models, and hyperparameters, all with static type checking.

DeepSeek-MoE is an innovative Mixture-of-Experts (MoE) language model featuring 16.4 billion parameters. It utilizes a unique architecture with fine-grained expert segmentation and shared experts isolation, allowing it to achieve performance comparable to DeepSeek 7B and LLaMA2 7B while requiring only about 40% of the computations. Trained from scratch on 2 trillion English and Chinese tokens, DeepSeek-MoE provides both base and chat model checkpoints for research and commercial use. It can be deployed on a single GPU with 40GB of memory without quantization, and offers quick start guides for installation and inference using Huggingface's Transformers. The project also provides scripts for fine-tuning the models on downstream tasks, supporting both DeepSpeed and QLoRA configurations.

DeepSeek-671B-SFT-Guide offers an open-source solution for the full parameter fine-tuning of DeepSeek-V3/R1 671B models. Developed by the Institute of Automation of the Chinese Academy of Sciences and Beijing Wenge Technology Co. Ltd., this guide includes comprehensive code and scripts covering the entire process from training to inference. It also shares practical experiences, common pitfalls, and solutions encountered during model training and deployment. Key features include implemented modeling files for DeepSeek-V3/R1 training logic, support for full parameter fine-tuning using data parallelism (DeepSpeed ZeRO) and sequence parallelism, and detailed instructions for environment setup, data preparation, training, model weight conversion, and inference deployment. The guide is designed for technical users looking to fine-tune large language models efficiently.

DiffSynth-Studio is an open-source Diffusion model engine developed and maintained by the ModelScope Community, focusing on aggressive technical exploration and academic use. It provides cutting-edge model capability support for diffusion models, including text-to-music, image editing, audio-video generation, and various training capabilities like LoRA and FP8. The platform aims to lower the threshold for technical exploration with detailed documentation and supports a wide range of models such as Stable Diffusion, SDXL, ACE-Step-1.5, JoyAI-Image, LTX-2.3, Anima, FLUX.2, and Qwen-Image. It also features advanced VRAM management and split training for optimized performance.

Nunchaku is a high-performance inference engine specifically designed for 4-bit neural networks, implementing the SVDQuant post-training quantization technique. This technology allows for 4-bit weights and activations while maintaining visual fidelity, as detailed in the accompanying ICLR 2025 Spotlight paper. The engine achieves significant memory reduction, up to 3.6x for 12B FLUX.1-dev models, and offers substantial speedups, such as 8.7x over 16-bit models on a 16GB laptop 4090 GPU by eliminating CPU offloading. Nunchaku also supports various features like LoRA, ControlNet, asynchronous offloading, and compatibility with ComfyUI, making it a versatile tool for accelerating diffusion models and other AI applications.

OSWorld is an open-source benchmark designed to evaluate multimodal AI agents performing open-ended tasks in real computer environments. It offers a robust framework for researchers and developers to test and compare the capabilities of their AI agents. The platform supports various virtualization technologies like VMware, VirtualBox, and Docker, with ongoing support for cloud platforms such as AWS. Key features include parallel execution of experiments, detailed result logging with screenshots and video recordings, and tools for manual task examination. OSWorld aims to standardize the benchmarking process for AI agents, providing clear metrics for success rates across different domains like Office, Daily, and Professional tasks.

filter-pruning-geometric-median is an open-source implementation of the Filter Pruning via Geometric Median method for accelerating deep convolutional neural networks. Developed in PyTorch, this tool enables researchers and developers to reduce the computational cost and memory footprint of their models without significant loss in accuracy. It supports both network-level and layer-level sparsity configurations, offering flexibility in how pruning is applied. The repository provides detailed usage instructions for integration with PyTorch and NNI, along with scripts for reproducing results on datasets like ImageNet and CIFAR-10, making it a valuable resource for model compression research and application.

Otter is an open-source multi-modal model developed by EvolvingLMMs-Lab, built upon the OpenFlamingo architecture. It excels in instruction-following and in-context learning, trained extensively on the MIMIC-IT dataset, which comprises 2.8 million interleaved image-text/video instruction-response pairs. Otter supports various tasks, including scene comprehension, reasoning, and multi-round conversations, and can process both image and video inputs. The project also introduces OtterHD for fine-grained interpretations of high-resolution visual input and MagnifierBench for evaluating tiny object recognition. It provides training scripts, pre-trained weights, and supports integration with Hugging Face models.

flux2 is the official inference repository for FLUX.2 models, offering state-of-the-art visual intelligence for image generation and editing. It provides minimal inference code to run these tasks with FLUX.2 open-weight models, including the FLUX.2 [klein] family, which boasts sub-second generation on consumer GPUs. The tool supports text-to-image generation, single-reference image editing, and multi-reference image editing. It is designed for developers and researchers, allowing for local installation and execution. The repository also details different model variants, their capabilities, licensing (Apache 2.0 for some, FLUX Non-Commercial License for others), and hardware requirements, making it suitable for various use cases from real-time applications to fine-tuning and research.

Opik, built by Comet, is an open-source platform designed to streamline the entire lifecycle of LLM applications, from prototype to production. It empowers developers to evaluate, test, monitor, and optimize their models and agentic systems with comprehensive tracing of LLM calls, conversation logging, and agent activity. Key features include advanced evaluation capabilities like LLM-as-a-judge for tasks such as hallucination detection and RAG assessment, experiment management, and integration into CI/CD pipelines. Opik also offers production-ready scalable monitoring dashboards, online evaluation rules, and dedicated SDKs for prompt and agent optimization, along with guardrails for safe AI practices. It supports a wide array of frameworks and offers client SDKs for Python, TypeScript, and Ruby.

Onepanel is an open-source, end-to-end computer vision platform designed to streamline the entire computer vision lifecycle. It provides a unified environment for labeling datasets, building models, training, tuning hyperparameters, deploying, and automating computer vision workflows. The platform is built to be flexible, supporting deployment on any cloud infrastructure as well as on-premises environments. By integrating various open-source projects like Argo, Couler, CVAT, JupyterLab, and NNI, Onepanel offers a comprehensive solution for machine learning and deep learning practitioners. It aims to simplify complex computer vision tasks from data preparation to model deployment and automation.

Prometheus-Eval is a comprehensive open-source repository designed for evaluating Large Language Models (LLMs) in various generation tasks. It leverages powerful models like Prometheus and GPT-4 to provide robust assessments. The tool supports multilingual meta-evaluation benchmarks, with recent iterations like M-Prometheus outperforming previous open LLM judges on multilingual meta-evaluation benchmarks such as MM-Eval and M-RewardBench. It also offers strong performance in English, surpassing Prometheus 2 7B and 8x7B on RewardBench. Prometheus-Eval facilitates both absolute grading, which assigns a score from 1 to 5, and relative grading, which compares two responses. It supports local inference via vllm and integration with LLM APIs through litellm, allowing users to utilize powerful evaluator LLMs like GPT-4.

PhiFlow is an open-source simulation toolkit designed for machine learning and optimization, primarily written in Python. It offers a differentiable PDE solving framework that seamlessly integrates with popular machine learning frameworks such as NumPy, PyTorch, Jax, and TensorFlow. This integration allows users to leverage automatic differentiation for building end-to-end differentiable functions that combine learning models with physics simulations. PhiFlow supports a wide range of applications, particularly in fluid dynamics, with features like built-in PDE operations, a flexible web interface for live visualizations, and object-oriented design for extensibility. It enables reusable simulation code across different backends and dimensionalities, making it a versatile tool for researchers and developers.

PyTorch-StudioGAN is a Pytorch library designed for implementing and researching Generative Adversarial Networks (GANs) for both conditional and unconditional image generation. It serves as a unified playground for machine learning researchers to compare and analyze new GAN ideas, offering implementations of 7 GAN architectures, 9 conditioning methods, 4 adversarial losses, and various regularization and augmentation modules. The library also provides an unprecedented-scale benchmark for generative models, including results from GANs, auto-regressive models, and Diffusion models. It supports multiple acceleration methods like distributed data-parallel training and mixed-precision training, ensuring flexibility and reproducibility in GAN research.

Python-ai-assistant, also known as Jarvis, is an open-source voice-commanding AI assistant built with Python 3.8. It offers a range of functionalities including speech recognition, text-to-speech interaction, and the execution of various commands. Users can interact with Jarvis via voice or text to perform tasks such as opening web pages, playing music, checking weather, setting alarms, and performing basic calculations. The assistant supports asynchronous command execution and allows for easy customization of voice commands and configurable assistant names. It also keeps a history of commands and learned skills in MongoDB, making it a versatile tool for personal automation.

PyTorch Frame is a modular deep learning framework built upon PyTorch, specifically designed for heterogeneous tabular data. It supports various column types including numerical, categorical, text, time, and images, enabling the creation of sophisticated neural network models. The library provides a flexible architecture for implementing existing and future deep learning methods, featuring state-of-the-art models, user-friendly mini-batch loaders, and benchmark datasets. It also facilitates integration with diverse model architectures, including Large Language Models, allowing users to encode text data with embeddings and train alongside other complex semantic types. PyTorch Frame aims to democratize deep learning research for tabular data, making it accessible for both novices and experts.

GPTZero is an open-source implementation of an AI model that determines if a given text was written by AI or a human being. The project aims to replicate the functionality of the original GPTZero, which gained significant attention for its ability to detect ChatGPT-generated content. This implementation leverages mathematical formulations and has been shown to produce results identical to the proprietary GPTZero.me. It is built using Python and relies on the Hugging Face transformers library, specifically utilizing models like Roberta for its underlying analysis. The tool can be used via a Python function or an interactive input script, making it accessible for developers and researchers to integrate into their own systems or use for direct text analysis.

rag-from-scratch is an open-source project designed to demystify Retrieval-Augmented Generation (RAG) by guiding developers through building it from scratch. It emphasizes local LLMs and avoids black boxes or cloud APIs, fostering a deep understanding of core RAG concepts. The project covers essential components such as embeddings, local vector database construction, retrieval strategies, and context-augmented generation. It offers step-by-step code walkthroughs, explaining every function and concept, making advanced AI approachable. Key learning areas include how embeddings work, building in-memory and LanceDB/Qdrant vector stores, basic and hybrid retrieval, query preprocessing, multi-query retrieval, and query rewriting. The project aims to provide a clear, practical, and comprehensive learning path for developers interested in RAG.

search is an open-source Go library designed for embedded vector search and semantic embeddings, utilizing llama.cpp. It offers an efficient solution for projects requiring semantic power without the complexities of traditional search systems. The library supports GGUF BERT models and provides GPU acceleration for quicker computations. It's particularly well-suited for datasets with fewer than 100,000 entries, offering features like llama.cpp integration without cgo, support for various BERT models in GGUF format, and precompiled binaries with Vulkan GPU support. Users can create and save search indexes from computed embeddings, enabling basic vector-based searches in Go applications.