Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 208 stars

YJiangcm/Lion

[EMNLP 2023] Lion: Adversarial Distillation of Proprietary Large Language Models

Overview: The Lion project, presented at EMNLP 2023, introduces an adversarial distillation framework that leverages a superior proprietary large language model (LLM) as a teacher to train a compact Student LLM. This process involves three stages: imitation, discrimination, and generation, aiming to enhance the Student LLM's performance. Key features: - The Teacher, Referee, and Generator roles are played by a compact Student LLM. - The model weights are released as delta weights to adhere to the LLaMA model license. - The Student LLM can achieve 95% capability of ChatGPT, demonstrating significant performance improvements. Tech stack: Python Why it matters: This project showcases the potential of adversarial distillation to enhance the capabilities of proprietary large language models, potentially leading to more efficient and effective AI systems.

Python 208 stars

ctrl-gaurav/effGen

[ICML 2026] effGen: Enabling Small Language Models as Capable Autonomous Agents

Overview: effGen is an open-source project designed to enable small language models to become capable autonomous agents, showcasing their capabilities through a series of benchmarks and demonstrations. Key features: - 📚 Prompt Library: A comprehensive library of prompts for various tasks. - ✨ Multimodal: Supports image, audio, and video inputs for more versatile interactions. - 🧑‍💻 Cookbook: Step-by-step guides for building AI agents. - 📊 Prometheus: Metrics and monitoring for better observability. - 🔭 OpenTelemetry: Tracing and monitoring with OpenTelemetry. - 🔔 SLOs & Alerting: Real-time monitoring and alerting for performance and issues. Tech stack: Python Why it matters: effGen demonstrates the potential of small language models to become autonomous agents, showcasing their capabilities through benchmarks and practical applications.

Python 208 stars

cvqluu/TDNN

Time delay neural network (TDNN) implementation in Pytorch using unfold method

### Overview: A simple implementation of the Time Delay Neural Network (TDNN) in PyTorch, utilizing the unfold method to process input sequences. This project is part of the x-vector network used in speech recognition. ### Key Features: - TDNN implementation using PyTorch. - Factorized TDNN (TDNN-F) implementation for improved efficiency. - Uses the unfold method to slide over input sequences. - Primarily designed for speech recognition applications. ### Tech Stack: PyTorch ### Why it Matters: This project provides a practical example of how to implement TDNN and Factorized TDNN in PyTorch, which can be useful for understanding and improving speech recognition models.

Python 208 stars

datamllab/OpenEBFM

Open-source code for training and scaling up energy-based foundation models.

Overview: Open-source code for training and scaling up energy-based foundation models. Key features: - Pretrained models for various datasets like ClimbMix, DCLM, and FineWeb. - Configuration for different training budgets. - Pretrained checkpoints available in the Hugging Face collection. - Zero-shot accuracy improves with model size and training tokens. Tech stack: Python Why it matters: This open-source code provides a framework for training and scaling up energy-based foundation models, making it useful for researchers and developers looking to improve language models and pretraining techniques.

Scala 208 stars

lightbend/kafka-with-akka-streams-kafka-streams-tutorial

Code samples for the Lightbend tutorial on writing microservices with Akka Streams, Kafka Streams, and Kafka

Overview: This tutorial provides an introduction to building Kafka-based microservices using Akka Streams and Kafka Streams. It compares Akka Streams and Kafka Streams, and demonstrates how to implement a stream processing application that also ingests updated parameters for a machine learning model and scores the data. Key features: - Kafka Streams integration - Akka Streams implementation - Stream processing application - Machine learning model scoring Tech stack: Scala, Akka Streams, Kafka Streams Why it matters: This tutorial helps developers understand the differences between Akka Streams and Kafka Streams, and provides a practical example of how to build a stream processing application that ingests data from Kafka and uses a machine learning model for scoring.

Kotlin 208 stars

llamatik/Llamatik

True on-device AI SDK for Kotlin Multiplatform (Android, iOS, Desktop, JVM, WASM). LLM inference, Speech-to-Text, Image Generation, Agent Framework and RAG — powered by llama.cpp, whisper.cpp and stable-diffusion.cpp.

Overview: Llamatik is a Kotlin-based True on-device AI SDK that provides a single Kotlin API for running AI applications across Android, iOS, Desktop, and WASM. It includes features such as LLM inference, Speech-to-Text, Image Generation, and Agent Framework powered by llama.cpp, whisper.cpp, and stable-diffusion.cpp. Key features: - LLM inference - Speech-to-Text - Image Generation - Agent Framework Tech stack: Kotlin Multiplatform Why it matters: Llamatik simplifies the development of AI applications by providing a unified Kotlin API for running AI on-device, ensuring offline-first and privacy-preserving capabilities. It supports multiple platforms and platforms, making it highly versatile for developers. Repository: llamatik/Llamatik GitHub stars: 207 GitHub description: True on-device AI SDK for Kotlin Multiplatform (Android, iOS, Desktop, JVM, WASM). LLM inference, Speech-to-Text, Image Generation, Agent Framework and RAG — powered by llama.cpp, whisper.cpp and stable-diffusion.cpp.

Python 208 stars

mage0535/hermes-memory-installer

🧠 Production-grade memory sidecar for AI agents — gbrain + Hindsight + 3-tier recall. Agent-agnostic, battle-tested. | 生产级外挂记忆系统，兼容Hermes/Claude/Cursor等任意AI智能体

### Overview: Memory Sidecar is a production-grade memory sidecar for AI agents, designed to be agent-agnostic and battle-tested. It integrates with Hermes, Claude, Code, and Cursor, ensuring seamless integration and improved recall across multiple sessions and knowledge sources. ### Key Features: - **Privacy-Safe Synthetic and Private Memory Evaluation Registries** - **Five Quality Metrics** - **Additive Governance Policy Metadata** - **Dry-Run-First gbrain Edge Planning** - **Canonical Runtime Path Resolution** - **Evaluation Trend Comparison** - **Private Registry Linting** - **Policy Validity/Conflict Metadata** - **Gbrain Plan Summaries** - **Operator Report** ### Tech Stack: Python ### Why It Matters: Memory Sidecar provides a robust and agent-agnostic solution for improving memory retention and recall, ensuring that AI agents can maintain context across multiple sessions and knowledge sources, enhancing their effectiveness and reliability.

C 207 stars

JustVugg/lumabri

Run large AI models across your computers and servers. Powered by Colibri. Pure C.

Overview: Lumabri is a private inference service that runs large AI models on your computers, powered by Colibri. It allows you to run multiple models concurrently without duplicating model weights, making it ideal for distributed computing setups. Key features: - Runs models across multiple computers - Shares model weights among workers - Tracks and manages memory usage - Supports multiple models in a single household - Provides a local workspace for monitoring Tech stack: Pure C Why it matters: Lumabri simplifies the process of running complex AI models on multiple computers, making it easier to scale and manage without the need for additional servers or dedicated hardware. Repository: JustVugg/lumabri GitHub stars: 204 GitHub description: Run large AI models across your computers and servers. Powered by Colibri. Pure C.

Python 207 stars

OpenMOSS/MOSS-Music

An open-source model for music captioning, lyrics transcription, structural analysis, and musical question answering

### Overview: MOSS-Music is an open-source model designed for music understanding, featuring specialized training on music-related tasks such as musical captioning, lyrics transcription, structural analysis, and long-form musical question answering. ### Key Features: - **Music Understanding Model**: MOSS-Music is built to unify music perception and understanding within a single model. - **Lyrics ASR & Time-aligned Transcription**: Accurate singing ASR with sentence- and word-level timestamps, robust to backing tracks. - **Musical Captioning & Tagging**: Natural-language descriptions of mood, genre, instrumentation, production style, and emotional trajectory. - **Key / Tempo / Chord Reasoning**: Identifies musical key, beats, downbeats, and chord progressions with timestamped chord transcription. - **Structural Analysis**: Segments a song into intro / verse / chorus / bridge / outro and reasons about repetition and contrast. ### Tech Stack: Python ### Why it Matters: MOSS-Music is designed to unify music perception and understanding within a single model, making it useful for tasks such as lyrics ASR, musical captioning, structural analysis, and more. ### Repository: OpenMOSS/MOSS-Music ### GitHub Stars: 206 ### Primary Language: Python ### Release: Two 8B models: MOSS-Music-8B-Instruct and MOSS-Music-8B-Thinking ### Contents: - **Introduction** - **Model Architecture** - **DeepStack Cross-Layer Feature Injection** - **Time-Aware

Python 207 stars

RobThePCGuy/Claude-Patent-Creator

USPTO patent creation system with MCP server + Claude Code plugin. Hybrid RAG search over MPEP/USC/CFR, BigQuery access to 76M+ patents, automated 35 USC 112 compliance checks, prior art search, diagram generation. GPU-accelerated with skills and autonomous agents.

### Overview: Claude Patent Creator is a patent creation and analysis system built with Claude Code, Python, and the Model Context Protocol (MCP) server. It leverages the MPEP, USC, and CFR for patent rules, BigQuery for patent data, and GPU-accelerated processing for efficiency. The system includes prior art searches, diagram generation, and automated compliance checks, making it accessible to anyone with a clear idea. ### Key features: - **AI-Powered Patent Rules**: Instantly get the actual rule with a citation for patent rule questions. - **Prior Art Search**: Finds existing patents to determine if your idea is new. - **Automated Compliance Checks**: Checks for 35 USC 112 compliance, ensuring your patent application is complete. - **Diagram Generation**: Creates patent-style drawings from a description or source code. - **Draft Review**: Guides you through the process, providing clean reports on mechanical issues. ### Tech stack: - Python 3.9+ - MCP Server (Python SDK 2.x) - PyTorch 2.9 ### Why it matters: - **Efficiency**: Automates the complex process of patent creation and analysis, saving time and reducing errors. - **Accessibility**: Open-source and accessible to anyone, regardless of legal or technical expertise. - **Accuracy**: Provides clear, actionable feedback and evidence, ensuring your patent application is strong and compliant. ### Repository: - **Repository**: RobThePCGuy/Claude-Patent-Creator - **Stars**: 206 - **Description

Python 207 stars

VITA-Group/Q-GaLore

Q-GaLore: Quantized GaLore with INT4 Projection and Layer-Adaptive Low-Rank Gradients.

### Overview: Q-GaLore is a memory-efficient training methodology that effectively enhances both pre-training and fine-tuning scenarios. It leverages low-rank gradients and lazy layer-wise subspace exploration to enable full-parameter learning while significantly reducing memory usage. This approach is particularly beneficial for training large language models like LLaMA-7B on limited memory resources. ### Key Features: - **Low-Precision Training**: Utilizes low-rank gradients to reduce memory usage. - **Lazy Layer-Wise Exploration**: Enables efficient exploration of subspaces without the need for full parameter updates. - **INT4 Projection**: Implements INT4 (Int8) projection for improved accuracy and reduced memory footprint. - **Layer-Adaptive Low-Rank Gradients**: Optimizes gradient computation for faster convergence. - **Lazy Layer-Wise Exploration**: Enables efficient exploration of subspaces without the need for full parameter updates. - **Quantization**: Incorporates quantization techniques to further reduce memory usage. - **Stochastic Round**: Uses stochastic rounding for improved numerical stability. - **Activation Checkpointing**: Utilizes activation checkpointing for efficient memory management. ### Tech Stack: - **Python**: The primary programming language used for implementing Q-GaLore. ### Why It Matters: Q-GaLore is a significant advancement in the field of memory-efficient training, particularly for large language models. By combining low-rank gradients and lazy layer-wise exploration, it enables full-parameter learning while requiring significantly less memory, making it ideal for training models on constrained hardware resources. This methodology is particularly valuable for applications where memory

Python 207 stars

Yi-Shi94/AMDM

Interactive Character Control with Auto-Regressive Motion Diffusion Models

Overview: This project introduces an Auto-Regressive Motion Diffusion Model (A-MDM) framework for interactive character control. The model allows for real-time motion generation and control, making it useful for applications requiring dynamic character animation. Key features: - Real-time motion generation - Interactive character control - Support for both training and inference - Includes real-time inpainting and reinforcement learning-based control Tech stack: The project utilizes PyTorch for the implementation. Why it matters: This framework provides a powerful tool for developers looking to create interactive and dynamic character animations, enhancing user experience in games, simulations, and other applications requiring real-time character control. Repository: Yi-Shi94/AMDM Primary language: Python GitHub stars: 207 GitHub description: Interactive Character Control with Auto-Regressive Motion Diffusion Models

Kotlin 207 stars

ferranpons/Llamatik

### Overview: Llamatik is a true on-device AI SDK for Kotlin Multiplatform, providing offline-first, privacy-preserving, and Kotlin Multiplatform support. It includes LLM inference, Speech-to-Text, Image Generation, Agent Framework, and RAG functionalities powered by llama.cpp, whisper.cpp, and stable-diffusion.cpp. ### Key Features: - **SDK Module (`com.llamatik:sdk`)**: All business logic extracted into a clean, UI-agnostic, publishable SDK. - **Assistant Facade**: A single, injectable facade for chat, RAG, image generation, model download, and session persistence. - **Agent Framework (`com.llamatik.sdk.framework`)**: A complete Kotlin DSL for building agentic AI applications, including agent creation, memory management, tool support, streaming, workflows, and observable events. ### Tech Stack: Kotlin Multiplatform - **LLM Inference**: Powered by llama.cpp - **Speech-to-Text**: Powered by whisper.cpp - **Image Generation**: Powered by stable-diffusion.cpp - **RAG / Vector Store**: Powered by llama.cpp ### Why It Matters: Llamatik simplifies the integration of AI capabilities into Kotlin Multiplatform projects, making it easier to build offline-first, privacy-preserving, and Kotlin Multiplatform applications. The SDK and Agent Framework provide a robust and flexible solution for building agentic AI applications in Kotlin.

Python 207 stars

flash131307/multi-agent-investment

Multi-agent equity research system with deterministic signal fusion — 3 specialized agents (ReAct, FinBERT, Plan-and-Solve + RAG) produce BUY/NEUTRAL/SELL decisions through a mathematical Decision Hub. No LLM in the critical path.

### Overview: A multi-agent equity research system that produces **BUY / NEUTRAL / SELL** decisions through a deterministic mathematical Decision Hub. No LLM is involved in the critical path, ensuring auditability and reproducibility. ### Key features: - **Deterministic decisions**: The Decision Hub is a pure mathematical function; same inputs always produce the same output. - **Three distinct reasoning patterns**: ReAct, FinBERT Funnel, Plan-and-Solve, each matched to its data modality. - **Sub-3s median latency**: Three agents execute in parallel with independent timeouts. - **Graceful degradation**: Produces calibrated decisions even when 1–2 agents fail. - **Hybrid RAG**: Reciprocal Rank Fusion over SEC 10-K filings, 0.82 recall@5. - **Domain-specific ML**: FinBERT replaces LLM for sentiment (50x cheaper, deterministic, F1=0.94 on financial text). - **294 tests**: across unit, integration, and E2E layers. ### Tech stack: Python ### Why it matters: This architecture addresses common issues with typical LLM approaches, such as prompt sensitivity, deterministic math, black-box reasoning, expensive sentiment inference, and single point of failure. The system ensures auditability and reproducibility by exposing intermediate values and maintaining 87% directional agreement even with 2-agent mode.

Python 207 stars

git-disl/PokeLLMon

### Overview: PokéLLMon is a local battle engine for Pokémon enthusiasts, designed to simulate battles between different Pokémon species. It leverages the power of Large Language Models (LLMs) to provide a dynamic and engaging experience. ### Key Features: - **Dynamic Pokémon Battles**: PokéLLMon allows users to engage in real-time battles, showcasing the strengths and weaknesses of various Pokémon. - **LLM Integration**: Utilizes Large Language Models to generate battle strategies and outcomes, adding an element of unpredictability and excitement. - **User-Friendly Interface**: Simplifies the setup process, making it accessible to both beginners and experienced Pokémon trainers. ### Tech Stack: - Python 3.8+ - OpenAI API v1.7.2 ### Why It Matters: PokéLLMon is a valuable tool for Pokémon enthusiasts, offering a unique way to experience and understand Pokémon battles. Its integration with OpenAI's API makes it a dynamic and exciting addition to the Pokémon community, enhancing the gaming experience with AI-driven strategies. ### Repository: git-disl/PokeLLMon ### Primary Language: Python ### GitHub Stars: 207 ### GitHub Description: No description provided. ### README Excerpt: No description provided.

Python 207 stars

junhsss/consistency-models

A Toolkit for OpenAI's Consistency Models.

### Overview: Consistency Models is a toolkit designed for the generation of high-quality images using generative models. It simplifies the process by enabling fast one-step generation while maintaining sample quality. ### Key Features: - **Fast One-Step Generation**: Achieves high sample quality without adversarial training. - **Few-Step Sampling**: Supports few-step sampling for trade-off between compute and sample quality. - **Self-contained Training Logic**: Includes all necessary schedules and loss functions. ### Tech Stack: Python ### Why It Matters: Consistency Models provide a straightforward and efficient way to generate high-quality images, making it easier for developers to experiment with generative models without the complexity of adversarial training. ### Repository: - **Repository**: junhsss/consistency-models - **GitHub Stars**: 207 - **GitHub Description**: A Toolkit for OpenAI's Consistency Models - **Primary Language**: Python ### Installation: ```sh $ pip install consistency ``` ### Quickstart: - **Basic Usage**: ```python from diffusers import DiffusionPipeline pipeline = DiffusionPipeline.from_pretrained( "consistency/cifar10-32-demo", custom_pipeline="consistency/pipeline", ) pipeline().images[0] # Super Fast Generation! 🤯 ``` - **Training**: ```python from pytorch_lightning import Trainer trainer = Trainer(max_epochs=8000, accelerator="auto") trainer.fit(cons

Python 207 stars

worldbench/LiDARCrafter

[AAAI 2026 Oral] LiDARCrafter: Dynamic 4D World Modeling from LiDAR Sequences

### Overview: LiDARCrafter is a unified framework designed for dynamic 4D world modeling from LiDAR sequences. It introduces the first 4D generative world model specifically tailored for LiDAR data, showcasing superior controllability and spatiotemporal consistency. The framework includes a tri-branch 4D layout conditioned pipeline that leverages language inputs to generate an editable 4D layout, guiding the synthesis of temporally stable LiDAR sequences. ### Key Features: 1. **4D Generative World Model**: A unique 4D generative world model specifically designed for LiDAR data. 2. **Tri-Branch 4D Layout Pipeline**: A pipeline that converts language inputs into an editable 4D layout, guiding the synthesis of temporally stable LiDAR sequences. 3. **Comprehensive Evaluation Suite**: A suite of metrics for evaluating LiDAR sequence generation, covering scene-level, object-level, and sequence-level metrics. 4. **Best Performance**: Demonstrated superior single-frame and sequence-level LiDAR point cloud generation performance on nuScenes, surpassing existing methods in foreground quality. ### Tech Stack: - Python ### Why It Matters: LiDARCrafter provides a comprehensive solution for generating dynamic 4D world models from LiDAR sequences, offering significant improvements in controllability, spatiotemporal consistency, and performance metrics. This framework is particularly useful for applications requiring dynamic and interactive 4D world modeling from LiDAR data. ### Repository: - **Repository**: worldbench/LiD

Python 206 stars

Corpus-OS/corpusos

Open-source protocol suite standardizing LLM, Vector, Graph, and Embedding infrastructure across LangChain, LlamaIndex, AutoGen, CrewAI, Semantic Kernel, and MCP. 3,330+ conformance tests. One protocol. Any framework. Any provider.

### Overview: Corpus OS is an open-source protocol suite that standardizes LLM, Vector, Graph, and Embedding infrastructure across LangChain, LlamaIndex, AutoGen, CrewAI, Semantic Kernel, and MCP. It offers a single protocol, error taxonomy, and metrics model, enabling interoperable AI frameworks and data backends across multiple providers. ### Key Features: - **Stable Protocols**: Standardized protocols across all four domains. - **Normalized Errors**: Consistent error handling with retry hints and machine-actionable scopes. - **SIEM-Safe Metrics**: Low-cardinality, tenant-hashed metrics for operational safety. - **Deadline Propagation**: Cancellation and cost control mechanisms. - **Vendor-Neutral**: Supports any provider, framework, and language. ### Tech Stack: Python ### Why It Matters: Corpus OS simplifies the integration of AI infrastructure by providing a single, standardized protocol, error handling, and metrics model. This standardization reduces the complexity and cost of cross-provider integration, making it easier to build and maintain applications that can be deployed across multiple providers without vendor lock-in. ### Repository: Corpus-OS/corpusos ### GitHub Stars: 205 ### GitHub Description: Open-source protocol suite standardizing LLM, Vector, Graph, and Embedding infrastructure across LangChain, LlamaIndex, AutoGen, CrewAI, Semantic Kernel, and MCP. 3,330+ conformance tests. One protocol. Any framework. Any provider.

Python 206 stars

FonaTech/Project_Chronos

⚡ Zero-Stall MoE Inference via Lookahead Prediction & Async DMA Prefetching. Optimized for SSD I/O with Hybrid MLA+Sliding Window Attention.

### Overview: Chronos is a storage-aware Multi-Headed Expert (MoE) inference framework optimized for SSD+DRAM hybrid inference, designed to eliminate the stall caused by offloading weights to RAM. ### Key features: - **Prefill-Time Loading**: Predicted experts are preloaded into memory before the first decode token, reducing the need for subsequent IO operations. - **Per-Expert Event Synchronization (M3)**: Expert events are recorded on the `H2D` stream, allowing the compute stream to wait only on the experts it actually needs, eliminating unnecessary synchronization. - **Hybrid MLAs and Sliding Window Attention**: Utilizes hybrid multi-layer attention (MLA) and sliding window attention for efficient inference. - **SSD+DRAM Hybrid Inference**: Designed to optimize performance on SSDs and DRAM-based hardware, reducing latency and improving overall inference speed. ### Tech stack: Python ### Why it matters: Chronos addresses the common issue of MoE offloading where models stall due to weight movement from SSD to RAM, ensuring smoother and more efficient inference on consumer hardware. By moving IO into prefill and synchronizing expert events at the expert-event level, Chronos maximizes pipeline efficiency and reduces latency, making it a significant improvement over traditional offloading methods. Repository: FonaTech/Project_Chronos GitHub stars: 206 Primary language: Python

Python 206 stars

RichardObi/medigan

medigan - A Python Library of Pretrained Generative Models for Medical Image Synthesis

Overview: `medigan` is a Python library designed to generate synthetic medical images, addressing the challenges of data scarcity and the lack of readily available generative models in medical imaging. Key features: - `medigan` provides user-friendly medical image synthesis capabilities. - It allows users to choose from a range of pretrained generative models. - The library supports data augmentation and domain adaptation techniques. - Synthetic datasets can be used for training or adapting AI models for clinical tasks. Tech stack: Python Why it matters: `medigan` offers a solution to the problem of data scarcity in medical imaging by providing access to pre-trained generative models, enabling the creation of synthetic datasets that can be used for training and evaluation of AI models in medical applications. Repository: RichardObi/medigan Primary language: Python GitHub stars: 206 GitHub description: medigan - A Python Library of Pretrained Generative Models for Medical Image Synthesis README excerpt:   ![medigan](docs/source/_static/medigan_logo.png) [![License](https://img.shields.io/github/license/RichardObi/medigan)](https://opensource.org/licenses/MIT) ![Continuous integration](https://github.com/RichardObi/medigan/actions/workflows/python-ci.yml/badge.svg) [![PyPI version](https://badge.fury.io/py/medigan.svg)](https://badge.fury.io/py/medigan) [![Conda version