You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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 + Flax - High-performance numerical computing with composable transformations (JIT, vmap, grad). Rising favorite for research and scientific ML.
Keras - High-level, beginner-friendly API that now runs on multiple backends (TensorFlow, JAX, PyTorch). Perfect for rapid experimentation.
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.
NLP & Transformers
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.
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.
Dask - Parallel computing for big data - scales pandas/NumPy/scikit-learn to clusters.
NumPy - Fundamental array computing library that powers almost every AI stack.
SciPy - Scientific computing algorithms (optimization, linear algebra, statistics, signal processing).
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.
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.
AutoKeras - Neural architecture search on top of Keras.
TPOT - Genetic programming-based AutoML for full pipeline optimization.
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).
Transformers - Library of pretrained transformer models and utilities for text, vision, audio, and multimodal training and inference.
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.
ONNX Runtime - High-performance inference and training for ONNX models across hardware.
🧠 2. Open Foundation Models
Pretrained language, multimodal, speech, and video models with publicly available weights.
Large Language Models (Base + Chat)
Qwen3.5 (Alibaba) - Native multimodal open series spanning from small to frontier-scale models (0.8B–122B). Released Feb 2026 with 397B total MoE params (17B active), strong in coding, math, vision, and instruction following.
DeepSeek-V3.2 / R1 (DeepSeek) - Mixture-of-Experts family with exceptional reasoning, math, and efficient large-scale inference.
Gemma 3 (Google) - Lightweight yet powerful open models with excellent efficiency for on-device use. Strong multilingual support across 100+ languages.
MiniMax-M2.1 / M1 (MiniMax) - Open-weight MiniMax model line spanning long-context reasoning and agentic software tasks, with strong tool use and publicly released weights for local deployment.
Kimi K2.5 (Moonshot AI) - Frontier open-weight MoE model with 256K context, strong coding and reasoning performance, and native multimodal + tool-use support for agentic workflows.
Mistral Large / Nemo / Small - High-performance model family with strong multilingual capability, tool use, and efficient deployment profiles.
Phi-4 / Phi-3.5 (Microsoft) - Small but highly capable models optimized for reasoning, edge devices, and on-device inference.
GLM-5 (Zhipu AI) - Strong open model line with solid coding, reasoning, and agentic-task performance.
OLMo 2 (Allen AI) - Fully open-source LLMs (1B–32B) with complete transparency: models, data, training code, and logs. Designed by scientists, for scientists.
Coding & Reasoning Models
DeepSeek-Coder-V2 / R1-Coder - Best-in-class open coding model (236B MoE). Outperforms closed models on many code benchmarks.
CodeLlama / CodeGemma - Meta's specialized coding variants built on Llama. Still heavily used for fine-tuning.
Qwen3-Coder-Next (Alibaba) - Leading open coding model. Strong Pareto frontier for cost-effective agent deployment.
StarCoder2 (BigCode) - 15B model trained on 600+ programming languages. Community favorite for transparency.
Multimodal Models (Vision + Language)
Qwen3-VL (Alibaba) - Latest flagship VLM with native 256K context (expandable to 1M), visual agent capabilities, 3D grounding, and superior multimodal reasoning. Major leap over Qwen2.5-VL.
InternVL3 (OpenGVLab) - Native multimodal pretraining with mixed preference optimization (MPO). Superior perception and reasoning over InternVL 2.5, extends to GUI agents and 3D vision.
GLM-4.5V / GLM-4.1V-Thinking (Zhipu AI) - Strong multimodal reasoning with scalable reinforcement learning. Compares favorably with Gemini-2.5-Flash on benchmarks.
LLaVA-OneVision - Successor to LLaVA 1.6 with expanded capabilities across vision-language tasks.
MiniCPM-V 2.6 - Handles images up to 1.8M pixels with top-tier OCR performance. Excellent for on-device deployment.
Gemma 3 (Google) - Lightweight multimodal supporting vision-language input, optimized for efficiency and on-device use.
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.
Ollama - Dead-simple local LLM runner with a one-line install, model registry, and OpenAI-compatible API.
MLX (Apple) - High-performance array framework + LLM inference optimized for Apple Silicon.
MLC-LLM - Deployment engine that compiles and runs LLMs across browsers, mobile devices, and local hardware.
Mastra - TypeScript-first agent framework with built-in RAG, workflows, tool integrations, observability and observational memory.
Mission Control - Cockpit for the agentic era — manage AI agent swarms with autonomous daemon, Field Ops for real-world execution, encrypted vault, and approval workflows.
Pi (badlogic) - Terminal coding agent with hash-anchored edits, LSP integration, subagents, MCP support, and package ecosystem.
Mistral-Vibe (Mistral) - Minimal CLI coding agent by Mistral. Lightweight, fast, and designed for local development workflows.
Nanocoder (Nano-Collective) - Beautiful local-first coding agent running in your terminal. Built for privacy and control with support for multiple AI providers via OpenRouter.
Domain-Specific Agents
Langflow - Visual low-code platform for agentic workflows.
SuperAGI - Dev-first autonomous AI agent platform.
Agent Memory & State
Letta (ex-MemGPT) - Platform for building stateful agents with advanced memory that learn and self-improve over time.
Mem0 - Universal memory layer for AI agents. Persistent, multi-session memory across models and environments.
Forgetful - MCP server for persistent AI agent memory. Stores atomic single-concept notes and auto-links them into a knowledge graph via semantic similarity. SQLite or PostgreSQL.
OpenCompass - Evaluation platform for benchmarking language and multimodal models across large benchmark suites.
SWE-rebench (Nebius) - Continuously updated benchmark with 21,000+ real-world SWE tasks for evaluating agentic LLMs. Decontaminated, mined from GitHub.
AutoTS - Automated time series forecasting with broad model selection, ensembling, anomaly detection, and holiday effects. Designed for production deployment with minimal setup.
llama.vim - Local LLM-powered code completion plugin for Vim/Neovim using llama.cpp. Fast, privacy-first, no API key needed.
CodeCompanion.nvim - AI-powered coding assistant for Neovim. Inline code generation, chat, actions, and tool use with support for multiple LLM providers.
- Dynamic computation graphs, Pythonic API, dominant in research and production. The current standard for most frontier AI work.
- End-to-end platform with excellent production deployment, TPU support, and large-scale serving tools.
+ Flax
- High-performance numerical computing with composable transformations (JIT, vmap, grad). Rising favorite for research and scientific ML.
- High-level, beginner-friendly API that now runs on multiple backends (TensorFlow, JAX, PyTorch). Perfect for rapid experimentation.
- Next-generation deep learning framework in Rust. Backend-agnostic with CPU, GPU, WebAssembly support.
- Minimalist ML framework for Rust. PyTorch-like API with focus on performance and simplicity.
- The de facto standard library for pretrained NLP models. 1M+ models, 250,000+ downloads/day. BERT, GPT, Llama, Qwen, and hundreds more.
- Classic library for sentence and image embeddings.
- Fast state-of-the-art tokenizers for training and inference.
- The gold standard for data analysis and manipulation in Python.
- Blazing-fast DataFrame library (Rust backend) - modern alternative to pandas for large-scale workloads.
- Parallel computing for big data - scales pandas/NumPy/scikit-learn to clusters.
- Fundamental array computing library that powers almost every AI stack.
- Scientific computing algorithms (optimization, linear algebra, statistics, signal processing).
- Industry-standard library for traditional machine learning (classification, regression, clustering, pipelines).
- Scalable, high-performance gradient boosting library. Still dominates Kaggle and tabular competitions.
- Microsoft's ultra-fast gradient boosting framework, optimized for speed and memory.
- Gradient boosting that handles categorical features natively with great out-of-the-box performance.
- Modern, define-by-run hyperparameter optimization with pruning and visualizations. Extremely popular in 2026.
- AWS AutoML toolkit for tabular, image, text, and multimodal data - state-of-the-art with almost zero code.
- Microsoft's fast & lightweight AutoML focused on efficiency and low compute.
- Neural architecture search on top of Keras.
- Genetic programming-based AutoML for full pipeline optimization.
- Simple API to make training scripts run on any hardware (multi-GPU, TPU, mixed precision) with minimal code changes.
- Microsoft's deep learning optimization library for extreme-scale training (ZeRO, offloading, MoE).
- Fast exact attention kernels that reduce memory usage and accelerate transformer training and inference.
- Optimized transformer building blocks and attention operators for PyTorch.
- High-level wrapper for PyTorch that removes boilerplate and adds best practices.
- High-performance inference and training for ONNX models across hardware.
- Native multimodal open series spanning from small to frontier-scale models (0.8B–122B). Released Feb 2026 with 397B total MoE params (17B active), strong in coding, math, vision, and instruction following.
- Mixture-of-Experts family with exceptional reasoning, math, and efficient large-scale inference.
- Lightweight yet powerful open models with excellent efficiency for on-device use. Strong multilingual support across 100+ languages.
- Open-weight MiniMax model line spanning long-context reasoning and agentic software tasks, with strong tool use and publicly released weights for local deployment.
- Frontier open-weight MoE model with 256K context, strong coding and reasoning performance, and native multimodal + tool-use support for agentic workflows.
- Small but highly capable models optimized for reasoning, edge devices, and on-device inference.
- Strong open model line with solid coding, reasoning, and agentic-task performance.
- Fully open-source LLMs (1B–32B) with complete transparency: models, data, training code, and logs. Designed by scientists, for scientists.
- Best-in-class open coding model (236B MoE). Outperforms closed models on many code benchmarks.
- Meta's specialized coding variants built on Llama. Still heavily used for fine-tuning.
- Leading open coding model. Strong Pareto frontier for cost-effective agent deployment.
- 15B model trained on 600+ programming languages. Community favorite for transparency.
- Latest flagship VLM with native 256K context (expandable to 1M), visual agent capabilities, 3D grounding, and superior multimodal reasoning. Major leap over Qwen2.5-VL.
- Native multimodal pretraining with mixed preference optimization (MPO). Superior perception and reasoning over InternVL 2.5, extends to GUI agents and 3D vision.
- Strong multimodal reasoning with scalable reinforcement learning. Compares favorably with Gemini-2.5-Flash on benchmarks.
- Successor to LLaVA 1.6 with expanded capabilities across vision-language tasks.
- Handles images up to 1.8M pixels with top-tier OCR performance. Excellent for on-device deployment.
- The gold-standard open speech-to-text model. Massive community fine-tunes available.
- High-quality open TTS with natural prosody and multilingual support.
- Zero-shot TTS with excellent voice cloning. Extremely popular in 2026.
- Open music and audio generation models.
- Fully open video generation model rivaling closed systems.
- High-quality open text-to-video model (5B-12B).
- 10B open video model with impressive motion and consistency.
- Motion module ecosystem for turning images into video.
- 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.
- Dead-simple local LLM runner with a one-line install, model registry, and OpenAI-compatible API.
(Apple) - High-performance array framework + LLM inference optimized for Apple Silicon.
- Deployment engine that compiles and runs LLMs across browsers, mobile devices, and local hardware.
- Official Python bindings for llama.cpp.
- User-friendly llama.cpp fork focused on role-playing and creative writing.
- Linux distribution for fully local AI inference on Raspberry Pi 5 and 4. Optimized for running open models at the edge.
- State-of-the-art serving engine with PagedAttention and continuous batching. Currently the fastest production-grade LLM server.
- Hugging Face's production-ready Rust-based server.
- Next-gen serving framework with RadixAttention.
- NVIDIA's official high-performance inference backend.
- vLLM fork optimized for role-play and creative writing.
- Kubernetes operator for LLM serving. GPU scheduling, model lifecycle management. Works with vLLM, SGLang, TensorRT-LLM.
- 8-bit and 4-bit optimizers + quantization.
- Activation-aware Weight Quantization toolkit.
- GPTQ quantization framework.
- Half-Quadratic Quantization - ultra-fast method rising in 2026.
- Highly optimized CUDA kernels for 4-bit/8-bit inference.
- Hardware-specific acceleration and quantization.
- Stateful, controllable agent orchestration.
- Role-based agent framework.
- Flexible multi-agent conversation framework.
- Framework for programming language model pipelines with modules, optimizers, and evaluation loops.
- SDK for building and orchestrating AI agents and workflows across multiple programming languages.
- Lightweight agent framework centered on tool use and code-executing workflows.
- Foundational library for agents, chains, and memory.
- Simulates an entire "AI software company".
- Lightweight multi-agent orchestration from OpenAI.
- Bleeding-edge enterprise multi-agent orchestration.
- Async-first multi-agent system.
- TypeScript-first agent framework with built-in RAG, workflows, tool integrations, observability and observational memory.
- Cockpit for the agentic era — manage AI agent swarms with autonomous daemon, Field Ops for real-world execution, encrypted vault, and approval workflows.
- Full-featured open-source AI software engineer.
- Extensible on-machine AI agent for development tasks.
- Terminal-native autonomous coding agent.
- Command-line pair-programming agent.
- Terminal coding agent with hash-anchored edits, LSP integration, subagents, MCP support, and package ecosystem.
- Minimal CLI coding agent by Mistral. Lightweight, fast, and designed for local development workflows.
- Beautiful local-first coding agent running in your terminal. Built for privacy and control with support for multiple AI providers via OpenRouter.
- Visual low-code platform for agentic workflows.
- Production-ready agentic workflow platform.
- Advanced multi-agent collaboration system.
- Dev-first autonomous AI agent platform.
- Platform for building stateful agents with advanced memory that learn and self-improve over time.
- Universal memory layer for AI agents. Persistent, multi-session memory across models and environments.
- MCP server for persistent AI agent memory. Stores atomic single-concept notes and auto-links them into a knowledge graph via semantic similarity. SQLite or PostgreSQL.
- Most popular open-source embedding database.
- High-performance vector search engine in Rust.
- GraphQL-native vector search engine.
- Scalable cloud-native vector database.
- Similarity search and clustering library for dense vectors with CPU and GPU implementations.
- Golang Low-latency graph + vector hybrid retrieval, Neo4j and qDrant driver compatible.
- Serverless vector DB optimized for multimodal data.
- PostgreSQL extension for vector similarity search.
- BAAI's best-in-class embedding family.
- High-performance text embeddings for retrieval.
- Open embedding model with long-context support and reproducible training.
- Full-featured RAG pipeline with advanced indexing.
- End-to-end NLP and RAG framework.
- Deep-document-understanding RAG engine.
- Knowledge-graph-based RAG.
- Golden RAG frontend with intuitive UI for retrieval and exploration.
- Advanced retrieval tools with late interaction models (ColBERT).
- Document processing toolkit for turning PDFs and other files into structured data for GenAI workflows.
- Best-in-class document preprocessing.
- Vision-language models for document retrieval.
- LLM-friendly web crawler that turns websites into clean Markdown for RAG and agentic workflows.
- Machine-first headless browser in Zig; rendering-free and ultra-lightweight for AI agent browsing.
- Node-based visual workflow editor for Stable Diffusion, FLUX, etc.
- Actively maintained Forge-based Stable Diffusion web UI with the familiar extension-driven workflow.
- Midjourney-style UI with beautiful out-of-the-box results.
- State-of-the-art open text-to-image model family.
- PyTorch library for diffusion pipelines spanning image, video, and audio generation.
- Full-featured creative studio.
- Latest open-weight diffusion model.
- Leading open Mixture-of-Experts text-to-video model.
- 13B-parameter systematic video generation framework. Leading quality among open models.
- First open-source infinite-length film generative model using AutoRegressive Diffusion-Forcing.
- Fast native 4K video generation.
- Official image-to-video and text-to-video implementation within Stability AI's generative models repository.
- Local-first music generation model with broad hardware support across Mac, AMD, Intel, and CUDA devices.
- Natural multilingual TTS with emotional control.
- Expressive zero-shot TTS.
- Gold standard for real-time voice cloning.
- State-of-the-art open image-to-3D and text-to-3D.
- Structured 3D latents for high-quality generation.
- Fast multi-view consistent 3D generation.
- Lightning-fast 3D reconstruction.
- End-to-end framework for training, rendering, and experimenting with NeRF and related 3D scene representations.
- High-performance 3D Gaussian Splatting library.
- One-stop unified framework for SFT, DPO, ORPO, KTO with web UI.
- YAML-driven full pipeline for SFT, DPO, GRPO.
- 2× faster, 70% less memory fine-tuning.
- Clean from-scratch implementations of 20+ LLMs.
- PyTorch-native library for post-training, fine-tuning, and experimentation with LLMs.
- Official library for RLHF, SFT, DPO, ORPO.
- Official library with LoRA, QLoRA, DoRA, etc.
- Ultra-fast custom kernels for training speedup.
- Advanced model merging tools.
- End-to-end pipeline for synthetic instruction data.
- High-performance data processing for LLM training.
- Open-source data labeling + synthetic data platform.
- High-fidelity tabular and relational synthetic data.
- Extreme-scale training optimizations.
- Unified system for 100B+ models.
- Distributed training framework and reference codebase for large transformer models at scale.
- Scalable distributed training.
- End-to-end open platform for the ML/LLM lifecycle.
- Git-like versioning for data and models.
- Open-source platform for experiment tracking, orchestration, data management, and model serving.
- Open-source tracing and experiment tracking.
- Unified framework to build, ship, and scale AI apps.
- Pipeline and orchestration framework for taking ML and LLM systems from development to production.
- Kubernetes-native ML/LLM platform.
- Kubernetes-based model serving.
- #1 open-source LLM observability platform.
- AI observability & evaluation platform.
- ML & LLM monitoring framework.
- Production-ready LLM evaluation platform.
- Programmable guardrails toolkit.
- Structure and validation for LLM outputs.
- Comprehensive input/output scanner.
- Real-time LLM hallucination guardrail with NLI + RAG fact-checking and token-level streaming halt.
- Open safety classifier models.
- LLM vulnerability scanner.
- LLM testing and red-teaming framework.
- De-facto standard for generative model evaluation.
- Holistic Evaluation of Language Models.
- Contamination-free coding benchmark.
- Hardened expert-level benchmarks.
- Evaluation platform for benchmarking language and multimodal models across large benchmark suites.
- The "Pytest for LLMs".
- End-to-end RAG evaluation framework.
- Evaluation toolkit for LLMs across multiple backends with reusable tasks, metrics, and result tracking.
- Standardized evaluation metrics.
- Largest open repository of datasets.
- Clean, reproducible Llama training data mix.
- Multimodal agent benchmark dataset.
- Safe reinforcement learning from human feedback.
- Complete recipes for full-stack alignment.
- High-performance distributed RLHF framework.
- Gold-standard for mechanistic interpretability.
- Scalable library for intervening on neural networks.
- Sparse autoencoders for interpretable features.
- PyTorch's official interpretability library.
- Automated LLM vulnerability scanner.
- Framework for custom adversarial testing.
- World's most widely used computer vision library.
- State-of-the-art real-time object detection.
- High-performance object detection library.
- Promptable image and video segmentation model with released checkpoints and training code.
- Differentiable computer vision library.
- Cross-platform multimodal pipelines.
- Production-ready RL algorithms.
- Single-file readable RL implementations.
- Multi-Agent Reinforcement Learning with JAX. Accelerated environments and baselines.
- GPU-accelerated robot learning framework.
- Standard RL environment API.
- Comprehensive benchmark for time-series models.
- Pretrained foundation models for time-series forecasting.
- Easy-to-use time-series forecasting library.
- Automated time series forecasting with broad model selection, ensembling, anomaly detection, and holiday effects. Designed for production deployment with minimal setup.
- PyTorch runtime and toolchain for deploying AI models on mobile, embedded, and edge devices.
- Intel's toolkit for edge deployment.
- Compiler stack for microcontrollers.
- Local-first personal AI assistant with multi-channel integrations and full agentic task execution.
- Most popular self-hosted ChatGPT-style interface.
- Web UI for running local LLMs with multiple backends, extensions, and model formats.
- Sleek modern chat UI.
- Feature-packed multi-LLM interface.
- Official open-source codebase for HuggingChat.
- Self-hostable personal AI assistant for search, chat, automation, and workflows over local and web data.
- GNOME/Linux desktop virtual assistant with integrated file editor, global hotkeys, and profile manager.
- All-in-one RAG + agents platform.
- Drag-and-drop LLM app builder.
- Privacy-first local desktop chatbot.
- Local-first AI app framework.
- Highly customizable role-playing frontend.
- Open-source AI coding autopilot for VS Code & JetBrains.
- Self-hosted AI coding assistant.
- Open-source IDE coding agent that can edit files, run commands, and use tools with user approval.
- Lets LLMs run code locally.
- Open-source editor-based coding agent with multiple modes and tool integrations.
- Local LLM-powered code completion plugin for Vim/Neovim using llama.cpp. Fast, privacy-first, no API key needed.
- AI-powered coding assistant for Neovim. Inline code generation, chat, actions, and tool use with support for multiple LLM providers.
- Chat and code generation inside notebooks.
- Legendary practical deep learning course.
- Massive collection of generative media workflows.
- Official tutorials: image classification, NLP, reinforcement learning.
- Run Transformers, Datasets, and more in Colab.
- Production-quality reference implementations.
