awesome-opensource-ai

Awesome Open Source AI

Curated open-source artificial intelligence models, libraries, infrastructure, and developer tools.

  svgdiagram.ai · Turn your notes into clean, editable SVG diagrams

---

Contributing

Contents

• 1. Core Frameworks & Libraries
• 2. Model Codebases & Model Families
• 3. Inference Engines & Serving
• 4. Agentic AI & Multi-Agent Systems
• 5. Retrieval-Augmented Generation (RAG) & Knowledge
• 6. Generative Media Tools
• 7. Training & Fine-tuning Ecosystem
• 8. MLOps / LLMOps & Production
• 9. Evaluation, Benchmarks & Datasets
• 10. AI Safety, Alignment & Interpretability
• 11. Specialized Domains
• 12. User Interfaces & Self-hosted Platforms
• 13. Developer Tools & Integrations
• 14. Resources & Learning

---

About this list

Awesome Open Source AI is a curated list of open-source projects for people building with AI.

The goal is to help readers find useful models, libraries, tools, infrastructure, datasets, and learning resources without sorting through a directory dump.

Projects do not need a minimum number of GitHub stars to be included. Stars can be useful context, but they are only one signal. A smaller project may belong here if it is useful, well-maintained, technically interesting, clearly documented, or important to a specific part of the AI ecosystem.

Good entries should have a clear reason to exist. They should help people build, study, run, evaluate, or understand AI systems.

---

1. Core Frameworks & Libraries

Core libraries and frameworks used to build, train, and run AI and machine learning systems.

Deep Learning Frameworks

• PyTorch - Dynamic computation graphs, Pythonic API, dominant in research and production. The current standard for most frontier AI work.
• TensorFlow - End-to-end platform with excellent production deployment, TPU support, and large-scale serving tools.
• JAX - High-performance numerical computing with composable transformations (JIT, vmap, grad). Rising favorite for research and scientific ML.
• Flax - Neural network library for JAX, designed for flexibility. Apache-2.0 licensed.
• dm-haiku - JAX-based neural network library from Google DeepMind. Elegant functional API with state management, widely used in DeepMind's research. Apache 2.0 licensed.
• Equinox - Elegant easy-to-use neural networks and scientific computing in JAX. Callable PyTrees with filtered transformations, seamless interoperability with the JAX ecosystem. Apache 2.0 licensed.
• Diffrax - Numerical differential equation solvers in JAX. Autodifferentiable and GPU-capable ODE/SDE/CDE solvers for scientific machine learning and neural differential equations. Apache 2.0 licensed.
• vit-pytorch - Comprehensive Vision Transformer (ViT) implementations in PyTorch. Reference implementations of all major vision transformer variants including ViT, DeiT, Swin, and more. MIT licensed.
• NumPyro - Probabilistic programming with NumPy powered by JAX for autograd and JIT compilation. Bayesian modeling and inference at scale.
• Keras - High-level, beginner-friendly API that now runs on multiple backends (TensorFlow, JAX, PyTorch). Perfect for rapid experimentation.
• tinygrad - Minimalist deep learning framework with tiny code footprint. The "you like PyTorch? you like micrograd? you love tinygrad!" philosophy - simple yet powerful.
• PaddlePaddle - Industrial deep learning platform from Baidu serving 23+ million developers and 760,000+ companies. China's first independent R&D framework with advanced distributed training and deployment capabilities.
• PyTorch Geometric - Library for deep learning on irregular input data such as graphs, point clouds, and manifolds. Part of the PyTorch ecosystem.
• timm (PyTorch Image Models) - The largest collection of PyTorch image encoders and backbones. 900+ pretrained models including ResNet, EfficientNet, Vision Transformer, ConvNeXt, and more with training and inference scripts. Apache 2.0 licensed.
• Triton - Language and compiler for writing highly efficient custom deep-learning primitives. Powers kernel optimizations in PyTorch, JAX, and other frameworks. MIT licensed.
• GGML - Tensor library for machine learning. The foundational C/C++ library powering llama.cpp and many on-device inference engines. MIT licensed.
• MLX - Array framework for machine learning on Apple silicon. Efficient unified memory design with NumPy-like API, automatic differentiation, and multi-device support. MIT licensed.
• notorch - Neural network framework in pure C with reverse-mode automatic differentiation, model training, and GGUF inference, without a Python runtime.

High-Performance Compute Libraries

• oneDNN - oneAPI Deep Neural Network Library. Cross-platform performance library of basic building blocks for deep learning, optimized for Intel CPUs, GPUs, and Arm architectures. Apache 2.0 licensed.
• ONNX - Open standard for machine learning interoperability. Open Neural Network Exchange provides an open ecosystem that empowers AI developers to choose the right tools as their project evolves. Apache 2.0 licensed.
• IREE - Retargetable MLIR-based machine learning compiler and runtime toolkit. Lowers ML models to unified IR that scales from datacenter to mobile and edge deployments. Apache 2.0 licensed.
• Modular Platform - Open-source AI compute and programming platform built around the MAX Engine and Mojo programming language.

Rust ML Frameworks

• Burn - Next-generation deep learning framework in Rust. Backend-agnostic with CPU, GPU, WebAssembly support.
• Candle (Hugging Face) - Minimalist ML framework for Rust. PyTorch-like API with focus on performance and simplicity.
• linfa - Comprehensive Rust ML toolkit with classical algorithms. scikit-learn equivalent for Rust with clustering, regression, and preprocessing.

Julia ML Frameworks

• Flux.jl - 100% pure-Julia ML stack with lightweight abstractions on top of native GPU and AD support. Elegant, hackable, and fully integrated with Julia's scientific computing ecosystem.
• MLJ.jl - Comprehensive Julia machine learning framework providing a unified interface to 200+ models with meta-algorithms for selection, tuning, and evaluation. MIT licensed.
• ModelingToolkit.jl - High-performance symbolic-numeric modeling framework for scientific machine learning. Automatically generates fast functions for model components like Jacobians and Hessians with automatic sparsification and parallelization. MIT licensed.

NLP & Transformers

• spaCy (Explosion AI) - Industrial-strength natural language processing with 75+ languages, transformer pipelines, and production-grade NER, parsing, and text classification.
• Transformers (Hugging Face) - The de facto standard library for pretrained NLP models. 1M+ models, 250,000+ downloads/day. BERT, GPT, Llama, Qwen, and hundreds more.
• sentence-transformers - Classic library for sentence and image embeddings.
• tokenizers (Hugging Face) - Fast state-of-the-art tokenizers for training and inference.
• fairseq2 - FAIR Sequence Modeling Toolkit 2. Complete rewrite of fairseq with modern PyTorch APIs, native support for LLM training (70B+ models), vLLM integration, and first-party recipes for instruction finetuning and preference optimization. MIT licensed.
• LibreTranslate - Self-hosted machine translation API powered by the Argos Translate engine. AGPL-3.0 licensed.

Data Processing & Manipulation

• Pandas - The gold standard for data analysis and manipulation in Python.
• Polars - Blazing-fast DataFrame library (Rust backend) - modern alternative to Pandas for large-scale workloads.
• cuDF - GPU DataFrame library from RAPIDS. Accelerates Pandas workflows on NVIDIA GPUs with zero code changes using cuDF.pandas accelerator mode.
• Dask - Parallel computing for big data - scales Pandas/NumPy/scikit-learn to clusters.
• DataFlow - LLM-ready data preparation system for turning raw PDFs, conversations, code, databases, and other sources into SFT, QA, and RAG-ready datasets.
• NumPy - Fundamental array computing library that powers almost every AI stack.
• SciPy - Scientific computing algorithms (optimization, linear algebra, statistics, signal processing).
• CuPy - NumPy and SciPy-compatible array library for GPU-accelerated computing in Python.
• NetworkX - Creation, manipulation, and study of complex networks. The foundational graph analysis library for Python data science.
• cuGraph - GPU graph analytics library with NetworkX-compatible API. 10-100x faster than CPU for large-scale graph algorithms. Apache 2.0 licensed.
• Datashader - High-performance large data visualization. Renders billions of points interactively without aggregation artifacts. BSD-3-Clause licensed.
• Zarr - Chunked, compressed, N-dimensional array storage. Scalable tensor data format optimized for cloud and parallel computing. MIT licensed.
• NVIDIA DALI - GPU-accelerated data loading and augmentation library with highly optimized building blocks for deep learning applications. Apache 2.0 licensed.
• Narwhals - Lightweight compatibility layer between DataFrame libraries. Write Polars-like code that works seamlessly across Pandas, Polars, cuDF, Modin, and more. MIT licensed.
• Ibis - Portable Python dataframe library with 20+ backends. Write Pandas-like code that runs locally with DuckDB or scales to production databases (BigQuery, Snowflake, PostgreSQL) by changing one line. Apache 2.0 licensed.
• skrub - Machine learning with dataframes for dirty categorical data. Preprocessing and feature engineering for heterogeneous data with seamless Pandas/Polars integration. BSD-3-Clause licensed.
• Oxen - Lightning fast data version control for machine learning. Optimized for large datasets with efficient diffing, branching, and collaboration. Apache 2.0 licensed.
• Pandera - Statistical data testing and validation for dataframes. Pydantic-like API for Pandas, Polars, and other dataframe libraries with type hints and lazy validation. MIT licensed.
• Snorkel - System for quickly generating training data with weak supervision. Programmatically label, build, and manage training data using labeling functions and probabilistic consensus models. Powers Snorkel Flow and used by Google, Apple, and Intel. Apache 2.0 licensed.
• DuckDB - High-performance analytical in-process SQL database system. Fast, reliable, portable, and easy to use with rich SQL dialect support. Perfect for data processing and analytics workloads. MIT licensed.
• FiftyOne - Visual AI development toolkit for visualizing, labeling, and evaluating visual datasets and models. Supercharges computer vision workflows with dataset exploration and model analysis. Apache 2.0 licensed.
• Label Studio - Multi-type data labeling and annotation tool with standardized output format. Configurable interface for images, text, audio, video, and time series with ML-assisted labeling. Apache 2.0 licensed.
• Delta Lake - Open-source storage framework enabling Lakehouse architecture with ACID transactions, scalable metadata handling, and unified batch/streaming processing. Apache 2.0 licensed.
• Apache Iceberg - High-performance open table format for huge analytic tables. Brings SQL table reliability to big data with time travel, hidden partitioning, and schema evolution. Works with Spark, Trino, Flink, Presto, Hive and Impala. Apache 2.0 licensed.
• Apache Hudi - Open data lakehouse platform for ingesting, indexing, storing, serving, transforming and managing data across cloud environments. Supports upserts, deletes and incremental processing on big data with built-in ingestion tools for Spark and Flink. Apache 2.0 licensed.
• lakeFS - Data version control for your data lake that transforms object storage into Git-like repositories. Enables atomic, versioned data lake operations with branching, committing, and merging for data pipelines. Apache 2.0 licensed.
• Apache Airflow - Platform to programmatically author, schedule, and monitor workflows. Industry-standard orchestration for data pipelines and ML workflows with 500+ integrations. Apache 2.0 licensed.
• Apache Spark - Unified analytics engine for large-scale data processing. In-memory cluster computing with high-level APIs in Python, Scala, Java, and R. Powers MLlib for distributed machine learning and Structured Streaming for real-time data. Apache 2.0 licensed.
• Apache Flink - Stream processing framework with powerful batch and streaming capabilities. High-throughput, low-latency runtime with exactly-once processing guarantees. Ideal for real-time AI inference pipelines and event-driven ML applications. Apache 2.0 licensed.
• Apache Beam - Unified programming model for batch and streaming data processing. Write pipelines once, run anywhere on Flink, Spark, or Google Cloud Dataflow. Portable, extensible, and enterprise-ready for AI data pipelines. Apache 2.0 licensed.
• Scrapy - Fast, high-level web crawling and scraping framework for Python. Extract structured data from websites at scale with built-in support for handling common challenges like pagination, cookies, and concurrent requests. BSD-3-Clause licensed.
• Temporal - Durable execution platform for reliable workflow orchestration. Build resilient data pipelines and ML workflows that survive failures and continue execution exactly where they left off. MIT licensed.
• Luigi - Python module for building complex pipelines of batch jobs. Handles dependency resolution, workflow management, visualization, and Hadoop integration. Built at Spotify and battle-tested in production. Apache 2.0 licensed.
• Mage.ai - Modern open-source data pipeline tool for integrating and transforming data. AI-native ETL/ELT platform with 100+ integrations, real-time monitoring, and collaborative features. Apache 2.0 licensed.
• Hamilton - Declarative dataflow framework for building testable, modular, self-documenting data pipelines. Encode lineage and metadata directly in Python functions. Originally from Stitch Fix, now Apache incubating. Apache 2.0 licensed.
• D-Tale - Visualizer for Pandas data structures with a Flask back-end and React front-end. Interactive data exploration with charting, filtering, and code export. LGPL-2.1 licensed.
• TextAttack - Python framework for adversarial attacks, data augmentation, and model training in NLP. Augment datasets to increase model robustness and generate adversarial examples. MIT licensed.
• uv - An extremely fast Python package and project manager, written in Rust. 10-100x faster than pip with built-in virtual environment management, dependency resolution, and lockfiles. Essential for modern AI/ML development workflows. Apache 2.0 and MIT dual-licensed.
• Vector - A high-performance observability data pipeline for collecting, transforming, and routing logs and metrics. Real-time data processing with 50+ sources and sinks including Kafka, S3, and Elasticsearch. Ideal for AI/ML log processing and data ingestion. MPL 2.0 licensed.

Classical ML & Gradient Boosting

• scikit-learn - Industry-standard library for traditional machine learning (classification, regression, clustering, pipelines).
• XGBoost - Scalable, high-performance gradient boosting library. Still dominates Kaggle and tabular competitions.
• LightGBM - Microsoft's ultra-fast gradient boosting framework, optimized for speed and memory.
• CatBoost - Gradient boosting that handles categorical features natively with great out-of-the-box performance.
• sktime - Unified framework for machine learning with time series. scikit-learn compatible API for forecasting, classification, clustering, and anomaly detection.
• StatsForecast - Lightning-fast statistical forecasting with ARIMA, ETS, CES, and Theta models. Optimized for high-performance time series workloads.
• MLForecast - Scalable machine learning for time series forecasting. Train any sklearn-compatible model on millions of time series with efficient feature engineering. Apache 2.0 licensed.
• cuML - GPU-accelerated machine learning algorithms with scikit-learn compatible API. 10-50x faster than CPU implementations for large datasets. Apache 2.0 licensed.
• SynapseML - Distributed machine learning on Apache Spark. Scalable, composable APIs for text analytics, vision, anomaly detection with seamless Python/Scala/R/.NET integration. MIT licensed.
• Darts - User-friendly forecasting and anomaly detection for time series. Unifies classical statistical models (ARIMA, ETS) with modern neural networks (N-BEATS, TFT, DeepAR) in a single scikit-learn compatible API. Apache 2.0 licensed.
• PyTorch Forecasting - Time series forecasting with PyTorch. Multiple neural architectures (N-BEATS, TFT, DeepAR) with in-built interpretation capabilities, built on PyTorch Lightning for distributed training. MIT licensed.

Data Engineering & Feature Stores

• DataHub - The #1 open-source metadata platform for data and AI. Data discovery, governance, and observability with 80+ connectors, column-level lineage, and AI assistant integration. Originally built at LinkedIn. Apache 2.0 licensed.
• OpenMetadata - Unified metadata platform for data discovery, observability, and governance. Column-level lineage, semantic search, and team collaboration with 70+ data service connectors. Apache 2.0 licensed.
• Amundsen - Data discovery and metadata engine from Lyft. PageRank-style search for data resources with usage-based ranking. LF AI & Data Foundation project. Apache 2.0 licensed.

Data Transformation & Analytics Engineering

• Apache Ossie - Vendor-neutral specification to standardize semantic models across analytics, BI, and AI agent platforms. Apache 2.0 licensed.
• dbt-core - Transform data using software engineering best practices. The industry-standard framework for analytics engineering with 15M+ monthly downloads. Enables version control, testing, and documentation for SQL transformations. Apache 2.0 licensed.
• SQLMesh - Scalable and efficient data transformation framework with dbt compatibility. Features automatic data lineage, time travel, and virtual data environments for testing. Optimized for large-scale data warehouses. Apache 2.0 licensed.
• SLayer - Semantic layer for AI-powered data analytics. Allows AI agents to describe data models and query the data using an expressive format with measures, dimensions, and filters, without writing raw SQL. MCP, CLI, API, and Python clients. Embeddable as a Python library. MIT licensed.
• WrenAI - Open-source Generative BI engine and context layer for AI agents to query databases and produce trusted SQL and dashboards. Apache 2.0 licensed.

Data Quality & Validation

• Deequ - Library built on top of Apache Spark for defining "unit tests for data". Measures data quality in large datasets with constraint verification, anomaly detection, and incremental validation. Used at Amazon for production data quality. Apache 2.0 licensed.
• Great Expectations - Always know what to expect from your data. Data validation, profiling, and documentation for data pipelines. Apache 2.0 licensed.
• ydata-profiling - One line of code for comprehensive data quality profiling and exploratory data analysis. Generates detailed reports for Pandas and Spark DataFrames including statistics, correlations, missing values, and data quality alerts. MIT licensed.
• Soda Core - Data contracts engine for the modern data stack. Define data quality checks in YAML and automatically validate schema and data across your pipelines. Supports 20+ data sources including Snowflake, BigQuery, and PostgreSQL. Apache 2.0 licensed.
• TFX (TensorFlow Extended) - End-to-end platform for deploying production ML pipelines. Data validation, transformation, model training, and serving with TensorFlow. Powers Google's production ML infrastructure. Apache 2.0 licensed.

Data Labeling & Annotation

• OpenRefine - Free, open-source power tool for working with messy data. Clean, transform, and extend data with web services. Formerly Google Refine. BSD-3-Clause licensed.

AutoML & Hyperparameter Optimization

• Optuna - Modern, define-by-run hyperparameter optimization with pruning and visualizations. Extremely popular in 2026.
• AutoGluon - AWS AutoML toolkit for tabular, image, text, and multimodal data - state-of-the-art with almost zero code.
• FLAML - Microsoft's fast & lightweight AutoML focused on efficiency and low compute.
• Katib (Kubeflow) - Kubernetes-native AutoML for hyperparameter tuning, early stopping, and neural architecture search. Framework-agnostic with support for TensorFlow, PyTorch, XGBoost, and custom training operators. Apache 2.0 licensed.

Interactive ML Apps & Notebooks

• Streamlit - The fastest way to build and share data apps. Transform Python scripts into beautiful web applications with minimal code. Widely used for ML model demos, data visualization, and internal tools.
• Gradio - Build and share delightful machine learning apps, all in Python. The de facto standard for creating interactive ML demos with automatic UI generation from function signatures. Powers thousands of Hugging Face Spaces.
• Marimo - A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.

Model Training & Optimization Utilities

• Hugging Face Accelerate - Simple API to make training scripts run on any hardware (multi-GPU, TPU, mixed precision) with minimal code changes.
• DeepSpeed - Microsoft's deep learning optimization library for extreme-scale training (ZeRO, offloading, MoE).
• FlashAttention - Fast exact attention kernels that reduce memory usage and accelerate transformer training and inference.
• xFormers - Optimized transformer building blocks and attention operators for PyTorch.
• PyTorch Lightning - High-level wrapper for PyTorch that removes boilerplate and adds best practices.
• fastai - Deep learning library providing practitioners with high-level components for state-of-the-art results. Built on PyTorch with a focus on usability and transfer learning. Apache 2.0 licensed.
• PyTorch Ignite - High-level library for training and evaluating neural networks in PyTorch with an engine, events & handlers system for maximum flexibility. BSD-3-Clause licensed.
• ONNX Runtime - High-performance inference and training for ONNX models across hardware.
• einops - Flexible, powerful tensor operations for readable and reliable code. Supports PyTorch, JAX, TensorFlow, NumPy, MLX.
• safetensors - Simple, safe way to store and distribute tensors. Fast, secure alternative to pickle for model serialization.
• torchmetrics - Machine learning metrics for distributed, scalable PyTorch applications. 80+ metrics with built-in distributed synchronization.
• torchao - PyTorch native quantization and sparsity for training and inference. Drop-in optimizations for production deployment.
• SHAP - Game theoretic approach to explain the output of any machine learning model. Industry standard for model interpretability.
• skorch - scikit-learn compatible neural network library that wraps PyTorch. Seamlessly integrate PyTorch models with scikit-learn pipelines, grid search, and cross-validation.
• NVIDIA Apex - PyTorch extension for mixed precision training and distributed training optimizations. Powers many production deep learning workloads with tools for automatic mixed precision (AMP), distributed data parallel, and fused optimizers. BSD-3-Clause licensed.

---

2. Model Codebases & Model Families

Canonical model-family repositories with useful code, recipes, evaluation tools, or engineering context. This is not a complete model leaderboard; use Hugging Face and model hubs for up-to-date weight discovery.

Language Model Families

• RWKV - Attention-free language model architecture with linear-time inference, training code, inference examples, and an active open-source ecosystem.
• MiniCPM - Compact open model family with practical code, deployment notes, and active edge/on-device focus.
• GPT-OSS - OpenAI open-weight model repository with inference examples, recipes, and deployment guidance.
• Mamba - State Space Model implementation with pretrained checkpoints, architecture code, and research tooling for efficient long-sequence modeling.
• GPT-NeoX - Large-scale language model training codebase from EleutherAI with distributed training support and historical open-model importance.
• GLM-5 - Open-source mixture-of-experts language model family optimized for long-horizon planning, agentic tasks, and coding. Apache 2.0 licensed.
• Bonsai Demo - Local runtime and model package for the Bonsai ternary reasoning model family, with GGUF and MLX support for on-device text, vision, and tool-calling workloads. Apache-2.0 licensed.

Multimodal & Vision-Language Codebases

• OpenCLIP - Open implementation of CLIP with training code, pretrained models, and zero-shot evaluation tooling.
• OmniParser - Vision-based GUI parsing model and tooling for computer-use agents.
• MiniCPM-V - Compact vision-language model family with edge-focused deployment examples and strong OCR-oriented use cases.
• Eagle - NVIDIA multimodal model codebase with open checkpoints and reusable research materials for vision-language and video-language tasks.
• NVIDIA Cosmos - Open platform of world models, tokenizers, and post-training tools designed for physical AI, robotics, and autonomous systems.

Speech & Audio Model Codebases

• Whisper - Canonical open speech-to-text model codebase with widespread ecosystem support and many downstream implementations.
• FunASR - Speech recognition toolkit with pretrained models, streaming support, diarization, VAD, and production-oriented examples.
• NVIDIA NeMo - Scalable framework and model codebase for speech, language, and multimodal AI with recipes and deployment guidance.
• Sherpa-ONNX - Complete speech toolkit with ASR, TTS, diarization, source separation, and VAD across embedded and edge environments via ONNX Runtime.
• MOSS-TTS - Open speech and sound generation family focused on expressive, long-form text-to-speech with streaming and multi-speaker support.
• VoxCPM - Open-sourced tokenizer-free multilingual speech synthesis model with high-quality TTS and style transfer workflows.
• VibeVoice - Open Frontier Voice AI toolkit spanning speech understanding, generation, and multilingual TTS workflows, with active research and deployment tooling.
• SpeechBrain - PyTorch speech toolkit with recipes for ASR, TTS, speaker recognition, and speech enhancement.
• Pocket TTS - Lightweight text-to-speech engine optimized for CPU inference with low latency and streaming support. MIT licensed.
• transcribe.cpp - C/C++ speech-to-text inference library running 16+ model families on the ggml runtime with GPU acceleration.
• Moonshine - Open-source on-device voice AI toolkit for low-latency speech-to-text, intent recognition, and text-to-speech.

---

3. Inference Engines & Serving

Inference runtimes, serving systems, and optimization tools for running models locally or in production.

Local / On-device Inference

• llama.cpp - Pure C/C++ inference engine with GGUF format support. The gold standard for CPU/GPU/Apple Silicon on-device running. Includes llama-server for OpenAI-compatible API. Now at 100K+ stars.
• Ollama - Dead-simple local LLM runner with a one-line install, model registry, and OpenAI-compatible API.
• Foundry Local - Open-source on-device AI platform covering discovery, model running, sandboxed execution, and evaluation of open models.
• Potato OS - Linux distribution for fully local AI inference on Raspberry Pi 5 and 4, optimized for running open models at the edge.
• MLC-LLM - Deployment engine that compiles and runs LLMs across browsers, mobile devices, and local hardware.
• WebLLM - High-performance in-browser LLM inference engine. Runs models directly in the browser with WebGPU acceleration.
• llama-cpp-python - Official Python bindings for llama.cpp.
• KoboldCpp - User-friendly llama.cpp fork focused on role-playing and creative writing.
• RamaLama - Container-centric tool for simplifying local AI model serving. Automatically detects GPUs, pulls optimized container images, and runs models securely in rootless containers with enterprise-grade isolation.
• LiteRT - Google's production-ready on-device ML and GenAI deployment framework. Supports Android, iOS, Web, Desktop, and IoT targets with GPU/NPU acceleration via a unified edge-first runtime. Apache 2.0 licensed.
• LiteRT-LM - Production-ready runtime for deploying LLMs on edge devices with low-latency inference and optimized hardware paths for mobile and embedded platforms.
• exo - Run frontier AI locally by connecting all your devices into an AI cluster. Features automatic device discovery, RDMA over Thunderbolt for 99% latency reduction, topology-aware auto parallel, and tensor parallelism. Uses MLX backend for distributed inference across Apple Silicon devices. Apache 2.0 licensed.
• ds4 - Native inference engine optimized for DeepSeek V4 and GLM models with Metal, CUDA, and ROCm support.
• qwen3.8-flash-next-in-c - Native C implementation of Qwen3.8-Flash-Next for local CPU inference, with terminal chat, a resident OpenAI-compatible API, and reproducible benchmarks.
• kimi-k3-in-c - Portable C99 inference engine for Kimi K3 that streams its 1.56 TB checkpoint from disk to run on CPU-only machines with as little as 8 GB RAM, with exact-output validation gates and reproducible measurements. Apache-2.0 licensed.
• omlx - Apple-centric inference server for local-first AI workflows with model management, GPU orchestration, and OpenAI-compatible APIs for self-hosted deployment. Apache 2.0 licensed.
• llmfit - Terminal tool and TUI that right-sizes LLM models to hardware specs and scores local compatibility across GPU, CPU, and RAM. MIT licensed.
• Needle - Compact 45M-parameter foundation model and 14MB inference engine for tool calling and structured extraction on tiny devices. MIT licensed.
• Colibri - Zero-dependency C inference engine that runs large Mixture-of-Experts models locally by streaming experts across disk, RAM, and VRAM.
• Claude Code Local - MLX-native server that speaks the Anthropic Messages API so the unmodified Claude Code CLI runs against local models on Apple Silicon, with parsing for local models' tool-call formats. MIT licensed.
• nemotron-omni-mlx - Pure MLX runtime for the vision and audio towers of NVIDIA Nemotron 3 Nano Omni on Apple Silicon, tested for parity against NVIDIA's PyTorch reference. MIT licensed.
• Magnitude - Hardware-aware local inference engine that profiles host hardware, recommends suitable open models, and tunes execution across Apple Silicon, NVIDIA, AMD, and CPU. Apache 2.0 licensed.
• llm-vram-calculator - Dependency-free TypeScript core that estimates inference memory (weights, KV cache, overhead) of any Hugging Face model from its config.json and safetensors metadata, covering MoE, MLA, sliding-window and linear-attention layers; runs in the browser at modelvram.com. MIT licensed.
• llmash - Ollama-compatible LLM server on a llama.cpp fork, tuned for speculative decoding on NVIDIA GPUs. Keeps Ollama's commands, API and model store, drafts with the model's own MTP head, and can uncensor a model with a control vector applied at load. MIT licensed.

High-performance Serving & API Servers

• llm-d - Kubernetes-native distributed LLM inference framework. Donated to CNCF by RedHat, Google, and IBM. Intelligent scheduling, KV-cache optimization, and state-of-the-art performance across accelerators.
• LMDeploy - Toolkit for compressing, deploying, and serving LLMs from OpenMMLab. 4-bit inference with 2.4x higher performance than FP16, distributed multi-model serving across machines.
• vLLM - State-of-the-art serving engine with PagedAttention and continuous batching. Currently the fastest production-grade LLM server.
• TokenSpeed - LLM inference engine for agentic workloads with a C++ scheduling control plane, Python execution plane, modular kernels, and KV-cache lifecycle management. MIT licensed.
• FreeToken - Edge-native Mixture-of-Experts serving engine for running frontier-scale open-weight models across consumer GPUs, CPUs, and host memory, with OpenAI- and Anthropic-compatible APIs. Apache-2.0 licensed.
• vLLM-Omni - Multi-modal inference stack extending vLLM for image, audio, and video generation workloads with a unified serving interface. Apache 2.0 licensed.
• LMCache - Supercharge LLM inference with the fastest KV Cache layer. 3-10x delay savings and GPU cycle reduction for multi-round QA and RAG. Integrates seamlessly with vLLM for distributed, high-throughput deployments. Apache 2.0 licensed.
• vLLM Production Stack - Kubernetes-native production stack for vLLM inference. Automated deployment, autoscaling, and monitoring for enterprise-grade LLM serving. Built by the vLLM team for seamless integration.
• Open Model Engine (OME) - Kubernetes operator for LLM serving with GPU scheduling and model lifecycle management across vLLM, SGLang, and TensorRT-LLM.
• SGLang - Next-gen serving framework with RadixAttention. Powers xAI's production workloads at 100K+ GPUs scale.
• TensorRT-LLM - NVIDIA's official high-performance inference backend.
• Aphrodite Engine - vLLM fork optimized for role-play and creative writing. Supports extensive quantization methods (AQLM, AWQ, GPTQ, GGUF, FP8) and modern samplers. Active development with multi-LoRA and speculative decoding support.
• AIBrix - Cost-efficient and pluggable infrastructure components for GenAI inference. Kubernetes-native control plane for vLLM with distributed KV cache, heterogeneous GPU serving, and intelligent routing. Apache 2.0 licensed.
• AISIX - Self-hosted Apache-2.0-licensed AI gateway written in Rust, with OpenAI-compatible endpoints, an Anthropic-compatible Messages endpoint, model routing, traffic policies, Prometheus metrics, and OTLP trace export.
• Triton Inference Server - NVIDIA's production-grade open-source inference serving software. Supports multiple frameworks (TensorRT, PyTorch, ONNX) with optimized cloud and edge deployment.
• mistral.rs - Fast, flexible Rust-native LLM inference engine built on Candle. Supports text, vision, audio, image generation, and embeddings with hardware-aware auto-tuning.
• KTransformers - Flexible framework for heterogeneous CPU-GPU LLM inference and fine-tuning. Enables running large MoE models by offloading experts to CPU with BF16/FP8 precision support.
• llamafile - Mozilla's single-file distributable LLM solution. Bundle model weights, inference engine, and runtime into one portable executable that runs on six OSes without installation.
• Xinference - Unified, production-ready inference API for LLMs, speech, and multimodal models. Drop-in GPT replacement with single-line code changes. Supports thousands of models with auto-batching and distributed inference.
• RTP-LLM (Alibaba) - Alibaba's high-performance LLM inference acceleration engine. Powers production LLM services across Taobao, Tmall, and Alibaba's international AI platform. Supports PagedAttention, FlashAttention, FlashDecoding, INT8/INT4 quantization, and heterogeneous hardware (GPU/ARM CPU/Intel). Apache 2.0 licensed.
• [LitServe (Lightning AI)](https://github.com/Light