Skip to content
BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

138 Commits
.gitignore
.gitignore
 
 
CONTRIBUTING.md
CONTRIBUTING.md
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 

Repository files navigation

Awesome Free Models Awesome

A curated list of free AI models, APIs, and tools you can use without paying a cent.

Last Updated Models Tools Sections License GitGem

✅ All links verified live on August 1, 2026. 330+ URLs checked. All links working (some sites block automated checks but work in browsers).

Running AI shouldn't require a credit card. This list curates genuinely free models — open-weight models you can self-host, free API tiers from major providers, and tools to run everything locally.

Contents

🧠 Open-Weight Models

📅 Last checked: August 1, 2026

Notable open-weight models you can download and run on your own hardware.

  • Llama 4 Scout / Maverick — Meta's latest MoE generation. Scout: 109B, 10M context. Maverick: 402B, 1M context. Native multimodal. [License]
  • DeepSeek V4 Pro — Apr 2026. 1.6T MoE (49B active). SWE-bench Verified 80.6% (top open-weight). 1M context. MIT license.
  • DeepSeek V4 — Core generation with extreme cost-efficiency. 1M context. MIT license.
  • DeepSeek-V4-Flash — Apr 2026. Efficiency-focused variant. 284B total (13B active). 1M context. MIT license.
  • Gemma 4 31B / 26B MoE / E4B / E2B — Fully permissive Apache 2.0. 256K context, native multimodal. New standard for open-weight.
  • Inkling (Thinking Machines Lab) — Jul 2026. 975B MoE (41B active). Leading US open-weight model. Native multimodal (text, image, audio). Apache 2.0. 1M context.
  • GLM-5.2 (Zhipu AI) — 744B MoE model optimized for autonomous coding and engineering tasks. 1M-token context. MIT license.
  • LongCat-2.0 (ByteDance) — Large-scale open-weight model for heavy agentic coding. MIT license.
  • MiniMax M3 — Frontier-tier 1M context, native multimodal + computer use. MSA architecture.
  • Trinity (Arcee AI) — 400B parameter enterprise model. Apache 2.0.
  • Step 3.7 Flash (StepFun) — May 2026. Apache 2.0. Native multimodal (image+video), strong agentic performance. Efficient enough for high-end local hardware.
  • Kimi K3 (Moonshot AI) — Jul 2026. 2.8T-parameter MoE (896 experts, ~50B active). World's largest open-weight model. 1M context, native vision + video. #1 Frontend Code Arena. Modified MIT license. Weights released Jul 27.
  • Kimi K2.6 (Moonshot AI) — Apr 2026. 1T-parameter MoE model. Modified MIT license. Exceptional coding (SWE-Bench ~54%) and multi-agent swarm orchestration.
  • Qwen 3.6-35B-A3B — Apr 2026. MoE variant with only 3B active parameters. Extremely efficient for consumer hardware. Apache 2.0.
  • InternLM 3 (Shanghai AI Lab) — Early 2026. Strong long-context reasoning and agentic performance. Competitive in open-weight benchmarks.
  • MiMo-V2.5-Pro (Xiaomi) — Apr 2026. 1.02T-parameter MoE (42B active). Optimized for complex agentic tasks, coding, and long-context.
  • Kimi K2.7 Code (Moonshot AI) — Jun 2026. 1T MoE specialized for long-running coding agents. +21.8% over K2.6 on coding benchmarks. Modified MIT.
  • Nemotron 3 Super (NVIDIA) — May 2026. 120B total (12B active). 1M context. Published weights, data, recipes, and eval infra. NVIDIA Open Model License.
  • Phi-4 14B (Microsoft) — 2025. Compact 14B dense model. Strong reasoning and code. MIT license. Excellent for on-device and small deployments.
  • Bonsai 8B (PrismML) — Apr 2026. Groundbreaking 1-bit quantized model. Extremely efficient for edge and consumer hardware (Apple Silicon).
  • Aether-7B-5Attn (VIDRAFT) — Jul 2026. 100% open foundation model (weights, data, code, logs). 7B MoE (~3B active) with heterogeneous attention. Apache 2.0.
  • Mistral Large 3 (Mistral) — Jun 2026. 675B MoE (41B active). European multilingual flagship. Frontier-class reasoning, native multimodal. Apache 2.0.
  • Mistral Small 3.1 (Mistral) — Mar 2025. Versatile 24B multimodal model. Strong text performance with native image understanding and 128K context. Apache 2.0.
  • Mistral Small 4 (Mistral) — Mar 2026. Hybrid MoE (6.5B active params) unifying instruction, reasoning, and multimodal capabilities. Efficient frontier-class model. Apache 2.0.
  • Command A+ (Cohere) — May 2026. Enterprise multimodal MoE optimized for sovereignty and multilingual RAG across 48 languages. Apache 2.0.
  • Apertus 1.5 (ETH Zurich / EPFL) — Jul 2026. Fully open LLM (weights, data, training code). 8B and 70B with image understanding, thinking mode, and tool use. Apache 2.0.
  • Hy3 (Tencent) — Jul 2026. 295B MoE (21B active). Strong reasoning and agentic performance. Competes with models 2-5x its size. Apache 2.0.
  • Hermes 4 (NousResearch) — Feb 2026. Self-improving agentic model with closed-loop learning. Curates own memory and builds skills from experience. Apache 2.0.
  • Snowflake Arctic — Apr 2024. Enterprise MoE model balancing high-quality performance with efficient training costs. Optimized for complex data operations. Apache 2.0.
  • Falcon 3 (TII) — Dec 2024. Compact high-performance model with strong reasoning. Designed for efficient deployment on resource-constrained hardware. TII Falcon-LLM License 2.0.
  • Apple OpenELM — Apr 2024. Family of efficient on-device SLMs using layer-wise attention scaling. Runs locally on Apple Silicon with full privacy. Apple Sample Code License.

🔌 Free API Providers

📅 Last checked: August 1, 2026

Providers offering free tiers to access models via API — no local hardware required.

  • Google AI Studio — Most generous free tier. Access Gemini 2.5 Flash, Gemini 2.0 Flash, and other models. Generous rate limits for prototyping.
  • OpenRouter — Aggregates 500+ models. Filter by "Free" to see models available at no cost. Includes experimental and subsidized open-weight models.
  • AnyAPI — 400+ models with OpenAI-compatible API. Free tier: 100K tokens/day, unlimited users. Includes free and basic models. No credit card required.
  • Groq — Ultra-fast inference. Free tier includes Llama, Gemma, Mixtral, Whisper models with generous daily rate limits.
  • Hugging Face Inference Providers — Free tier for thousands of community models. Rate-limited but excellent for testing.
  • NVIDIA NIM — Free API access to accelerated versions of Llama, Mistral, Gemma, and more on NVIDIA infrastructure.
  • DeepInfra — Serverless inference. Free tier with daily rate limits for popular open-source models.
  • Together AI — Free trial credits for new users. Fast inference on open-source models.
  • Fireworks AI — Free tier for community models. Optimized for low latency.
  • SiliconFlow — Rising platform with free access to many open-source models.
  • Cloudflare Workers AI — Free tier for running select open-source models at the edge.
  • Black Forest Labs — Free Flux 2 Dev and Flux Kontext Dev image generation via API. Rate-limited, no credit card required.
  • Replicate — Free tier with limited credits for running open-source models.
  • Poe (Quora) — Free tier with daily credits for GPT-4 mini, Claude instant, and community bots.
  • Qwen Studio (Alibaba) — Free access to Qwen 3.6-Plus, Qwen 3.6-Max, and other Qwen models via web chat and API. 1M token context for agentic coding.
  • Ollama Cloud — Free tier for running open-source models on Ollama's cloud infrastructure. Light usage with session limits (reset every 5 hours) and weekly limits. 1 concurrent model. Same ollama run command as local. Prompt/response data never logged or trained on.
  • Mistral AI (La Plateforme) — Free API tier with access to Mistral Large, Mistral Nemo, Codestral and more. 1 req/s, 500k tokens/min. Requires phone verification and data usage opt-in.
  • Cohere — Free evaluation API key for Command R, Command R+, Embed, and Rerank models. 20 req/min, 1,000 req/month.
  • DeepSeek Platform — Free API credits for new users (5M tokens). Access to DeepSeek V4, DeepSeek-R1, and other models. Generous free allocation.
  • GitHub Models — Free tier for GitHub users. Access GPT-4o, Llama 3.3, Mistral, and more with rate-limited playground and API.
  • Hyperbolic — Open-access AI cloud with affordable inference. Free compute credits via referral program. Supports Llama, Qwen, DeepSeek, and other open models.
  • Novita AI — Free credits for testing 100+ models including Llama, Qwen, DeepSeek, and Mistral. OpenAI-compatible API with competitive pricing beyond the free tier.
  • Anakin.ai — 30 daily free credits for accessing multiple AI models. Web chat interface and API access. Supports GPT-4, Claude, and open-weight models.
  • Nebius AI — $100 free credits for new users. AI Studio with access to Llama, Qwen, DeepSeek, and other open-weight models. Fast inference on NVIDIA H100 infrastructure.
  • Fal.ai — Free starter credits for AI inference. Fast, serverless platform supporting Llama, Flux, and Stable Diffusion models. Pay-as-you-go beyond free tier.
  • Vercel AI Gateway — $5/month free credits for the AI Gateway. Proxy and cache requests across multiple LLM providers. SDK is open-source and free.
  • AI21 Labs — $10 trial credits for accessing Jamba 1.5, Jamba 1.6, and other AI21 models. Valid for 3 months. Requires account sign-up.
  • Amazon Bedrock — $200 AWS credits for new customers. Access to Llama, Mistral, Claude, Titan, and other foundation models via API.
  • Microsoft Foundry (Azure) — $200 free trial credits (30 days). Access GPT-4o, Llama, Mistral, Phi, and other models via Azure's unified AI platform.
  • RunPod — Free credits for serverless GPU inference. Deploy open-weight models as serverless endpoints. Supports Llama, Qwen, DeepSeek, and more.
  • Cerebras — Free tier with ultra-fast inference on Llama, Gemma, and Mistral models. No credit card required. Note: Model availability fluctuates.
  • BazaarLink — Free OpenAI-compatible API with auto:free routing to zero-cost models. No credit card, no trial expiry. 10 RPM, 130 req/day.
  • Kimi API (Moonshot) — Free tier for new accounts with access to Kimi K2.5 (128K context). Also free via NVIDIA NIM. OpenAI-compatible.
  • Alibaba DashScope — Free tier for Qwen models. 1M tokens/month. OpenAI-compatible API.
  • SambaNova Cloud — Free tier with $5 credits (30-day). Fast RDU inference for Llama 3.1 405B, Llama 3.3 70B, DeepSeek V3.1/V3.2, Qwen 2.5, gpt-oss-120b. 20 RPM, 200K tokens/day. No credit card required.
  • OVHcloud AI Endpoints — EU-hosted, GDPR-compliant free tier. No registration required for anonymous tier. Models: Qwen, Mistral, Llama, DeepSeek, gpt-oss-120B, embeddings, image generation. 12 RPM. OpenAI-compatible.
  • Chutes.ai — Community-powered free GPU inference for open-source models. DeepSeek-R1, Llama 3.1 70B, Qwen 2.5 72B. OpenAI-compatible API. No credit card.
  • ModelScope — Chinese platform with 50+ free open models. Qwen, DeepSeek, GLM, and more. No credit card required.
  • Z.ai (Zhipu AI) — Free tier for GLM models including GLM-4.5, GLM-4V. No credit card required.
  • LongCat AI — Free API for LongCat open-weight models. One-time 10M token grant after signup + KYC. Free cached tokens. OpenAI-compatible. MIT license.
  • Coze (ByteDance) — Free bot-building platform with API access to GPT-4o, Gemini 1.5 Pro, and other models. No credit card required. OpenAI-compatible.
  • Free.ai — 400+ AI tools via a single OpenAI-compatible API. Free tier: 30K tokens/day, no credit card. Chat, image, video, music, voice, OCR, translation.
  • Requesty — Free AI API with 200 requests/day. Works with Claude Code, Cline, Cursor. No credit card. OpenAI-compatible.
  • AINative Studio — 84+ models (Llama, DeepSeek, Mistral, Qwen). Free tier: 10M tokens/month. No credit card required. 60 RPM.
  • CloudCode.ONE — Free tier for coding agents. Powered by GLM-4.7-Flash. OpenAI and Anthropic-compatible API. No credit card required.
  • ZeroLimitAI — Free OpenAI-compatible API with model: "auto" routing to the best free model. No credit card, no trial expiry. Lifetime free tier available.
  • Chat Oripe — 2M free tokens/month. OpenAI-compatible API with GPT-4 and Claude access. No credit card required.
  • FreeTheAi (da-jb) — Open-source. Free AI API via Discord signup. No daily cap, 30 RPM. OpenAI-compatible, image and video generation.
  • OpenCode Zen — Curated AI gateway with 7 free models (DeepSeek V4 Flash Free, MiMo-V2.5 Free, Nemotron 3 Ultra Free, Big Pickle, Qwen 3.6 Plus Free, MiniMax M3 Free, North Mini Code Free). OpenAI-compatible API. No credit card required.

🖼️ Image & Video Generation

📅 Last checked: August 1, 2026

Free, open-weight image and video generation models — run locally or via free APIs.

  • FLUX.2-dev (Black Forest Labs) — 2.6K★. 32B rectified flow transformer. SOTA open T2I, single/multi-reference editing, in/out-painting. Updated VAE. FLUX.1-dev Non-Commercial License.
  • ERNIE-Image / ERNIE-Image-Turbo (Baidu) — 0.5K★. 8B DiT SOTA among open-weight models. Strong text rendering, layout control. Turbo: 8-step generation. Apache 2.0.
  • Z-Image (Tongyi Lab / Alibaba) — 11.8K★. Open-weight T2I with strong GenEval scores. Z-Image-Turbo for 4-step generation. Apache 2.0.
  • Pollinations.ai — Free image generation API. No API key or signup needed. Text-to-image, image-to-image. OpenAI-compatible. Integrates with ComfyUI.
  • OpenImageGen (Hugging Face) — Free, open-source image generation playground. Supports multiple community models via diffusers. Apache 2.0. Note: Requires Hugging Face login.
  • ComfyUI — 123K★. Node-based image and video generation UI. Run FLUX, Stable Diffusion, and more locally. GPL-3.0.

🔀 Free API Routers

📅 Last checked: August 1, 2026

Open-source tools that route requests across multiple AI providers — unified API, automatic failover, and cost optimization.

  • 9Router — Open-source gateway connecting 40+ providers with RTK token compression (2-4x reduction). One API key for all services. MIT license. GitHub
  • OmniRoute — Full-stack AI gateway with 250+ providers, 90+ free. TypeScript, runs on Web/Desktop/Android. Prompt compression, 3-level proxy for geo restrictions. GitHub
  • LiteLLM — Python-based proxy unifying 100+ LLMs behind a single API. Spend tracking, virtual keys, production-ready. MIT license. GitHub
  • Portkey AI Gateway — Production guardrails and routing for AI apps. Hybrid open-source (community) and managed (enterprise) tiers. GitHub

💻 Local Inference Tools

📅 Last checked: August 1, 2026

Run models on your own machine — no API keys needed, full privacy.

  • Ollama — The easiest way to run local LLMs. One command to download and run any model. macOS, Linux, Windows. GitHub
  • LM Studio — Polished desktop GUI. Browse, download, and chat with models. Built-in model browser and local API server.
  • llama.cpp — High-performance C++ inference engine. Runs on CPU and GPU. Supports GGUF quantization. Powers most other local tools.
  • Jan — Open-source ChatGPT alternative for desktop. Built-in model downloader, local API server. GitHub
  • GPT4All — ⚠️ Unmaintained since Feb 2025. Privacy-focused local chatbot. Runs on consumer hardware. Built-in model browser. GitHub
  • text-generation-webui (Oobabooga) — Feature-rich web UI. Supports multiple backends (Transformers, llama.cpp, ExLlama, AutoGPTQ).
  • LocalAI — Drop-in OpenAI API replacement. Run models locally with an OpenAI-compatible API. GitHub
  • KoboldCPP — Single-file executable for running GGUF models. Focused on story generation but general-purpose.
  • llamafile (Mozilla) — Distributable single-file executables that run LLMs. No installation needed.
  • vLLM — High-throughput production inference engine. Uses PagedAttention for efficient serving.
  • SGLang — Fast inference framework with structured generation and RadixAttention.
  • TensorRT-LLM (NVIDIA) — NVIDIA's optimized inference engine. Best performance on NVIDIA GPUs.
  • ExLlamaV3 — Optimized inference for Llama-family models. Successor to ExLlamaV2. Fastest option for single-GPU inference.
  • Aphrodite Engine — High-performance LLM serving engine with advanced quantization support.
  • TabbyAPI — Lightweight, fast OpenAI-compatible API server for ExLlamaV2.
  • LlamaEdge — Lightweight inference framework for edge devices. OpenAI-compatible API for open-source models. Runs on WasmEdge for portability. GitHub
  • MLC LLM — Universal deployment engine by UW/SJTU. Runs LLMs on any hardware — laptops, phones, browsers. OpenAI-compatible API.
  • WebLLM — In-browser LLM inference via WebGPU. Runs models directly in your browser with zero setup. No server needed.
  • FastChat (LMSYS) — Open platform for training, serving, and evaluating LLMs. Provides OpenAI-compatible API and web UI for local models.
  • Hugging Face TGI — 10.9K★. Production-grade serving toolkit for large language models. Note: Archived by Hugging Face (Mar 2026). Consider vLLM, SGLang, or TGI forks for active development.
  • DeepSpeed (Microsoft) — Deep learning optimization library with inference acceleration. Enables running larger models on limited hardware through ZeRO optimization.
  • AirLLM — Run large models (70B+) on consumer hardware with limited memory. Loads models layer-by-layer for extreme memory efficiency. Actively maintained (last push Jul 2026).
  • AI Toolkit for VS Code (Microsoft) — VS Code extension to browse, test, fine-tune, and deploy models locally. Integrates ONNX and llama.cpp.
  • Ollama Grid Search — Desktop utility for systematic model evaluation. Test multiple models, prompts, and inference parameters side-by-side via a Rust/React GUI.
  • oMLX — 18.2K★. LLM inference server for Apple Silicon with continuous batching, tiered KV caching (hot RAM + cold SSD), and macOS menu bar app. OpenAI and Anthropic compatible. Apache 2.0.
  • MTPLX — 1.1K★. Native MTP speculative decoding on Apple Silicon — ~2x faster decode with no external drafter. Mac app + CLI, OpenAI/Anthropic compatible server. Auto-tunes draft depth per machine. Apache 2.0.

💬 AI Chatbot UIs

📅 Last checked: August 1, 2026

Free, open-source web interfaces for chatting with AI models — self-host or use hosted versions.

  • Open WebUI — Feature-rich ChatGPT-like interface for Ollama and OpenAI-compatible backends. RAG, image generation, multi-user. GitHub
  • LibreChat — Open-source ChatGPT clone supporting 40+ providers, multi-user, plugins, and RAG. Note: Company acquired by ClickHouse; repo actively maintained (last push Aug 2026). GitHub
  • AnythingLLM — All-in-one desktop app for chatting with documents and models. Built-in RAG pipeline. GitHub
  • Big-AGI — Feature-rich AI chat with personas, multi-model support, voice, and code execution. GitHub
  • Lobe Chat — Multi-agent orchestration platform with plugin system and multi-provider support. GitHub

🎵 Audio & Speech Models

📅 Last checked: August 1, 2026

Free, open-weight text-to-speech (TTS), speech-to-text (STT), and voice generation models you can run locally.

  • Qwen3-TTS (Alibaba) — 12.5K★. Voice cloning, voice design, 10 languages. Streaming support with 97ms TTFB. 0.6B/1.7B. Apache 2.0.
  • Chatterbox (Resemble AI) — 25.5K★. SOTA open-source TTS. Multilingual V3 (23+ languages, 0.5B). Turbo: 350M for low-latency agents. Paralinguistic tags. MIT.
  • MOSS-TTS Family (MOSI.AI/OpenMOSS) — 3.9K★. 8B flagship + 100M Nano (CPU). Voice cloning, dialogue generation, sound effects, realtime streaming. Apache 2.0.
  • Orpheus-TTS (Canopy Labs) — 6.2K★. Llama-3b backbone, human-like speech, zero-shot voice cloning, emotion tags. ~200ms streaming latency. Apache 2.0.
  • NeuTTS (Neuphonic) — 6K★. On-device TTS with instant voice cloning. GGUF quantized for CPU/mobile. 120M Nano and 360M Air variants. Apache 2.0.
  • Faster-Whisper — 25K★. CTranslate2-based Whisper for 4x faster transcription. MIT.

🤖 AI Coding Assistants

📅 Last checked: August 1, 2026

Free tools that integrate AI into your development workflow.

  • Continue.dev — Company acquired by Cursor; repo actively maintained (last push Aug 2026). Open-source AI code assistant for VS Code and JetBrains. GitHub
  • Aider — AI pair programming in the terminal. Edits code in your local git repo. Supports GPT, Claude, and local models. GitHub
  • Gemini CLI (Google) — Jul 2026. Open-source terminal agent with generous free Gemini quota. Supports agentic coding workflows.
  • Kilo Code — 2026. VS Code/JetBrains agentic coding extension with model-agnostic support and Plan/Act oversight.
  • Tabby — Self-hosted AI coding assistant with no dependency on external services. GitHub
  • Cody (Sourcegraph) — Free tier for individuals. Chat, autocomplete, and commands with codebase context.
  • Llama Coder (Nutlope) — Free AI code generation tool. Generate entire apps from prompts.
  • Bolt.new (StackBlitz) — Free tier for AI-powered full-stack web app development in browser.
  • Claude Code (Anthropic) — Terminal-based AI coding assistant. Most features require a Claude subscription or API credits. Limited free usage via terminal CLI.
  • Cursor 3 — Apr 2026. AI-native code editor with deep model integration and agentic features. Free tier available.
  • CodeBuff — CLI-based AI coding assistant that understands entire codebases. Multi-agent architecture, works with any model provider through natural language instructions.
  • Pi — Open-source terminal AI coding agent with a unified multi-provider API. Model-agnostic, supports OpenAI, Anthropic, Google, and any OpenAI-compatible endpoint. Extensible plugin architecture. GitHub
  • Cline — Popular autonomous VS Code agent. Creates/edits files, runs terminal commands, browses web. Open-source, BYOK (bring your own API key). GitHub
  • OpenHands — Autonomous AI software engineer. Navigates file systems, runs shell commands, tests code in browser. Self-hostable. GitHub
  • Goose — Open-source CLI agent for complex software engineering tasks. Extensible plugin system. Built by Block/Square. GitHub
  • Qwen Code — 2025. Open-source terminal AI coding agent with 26K+ stars. Multi-protocol (OpenAI, Anthropic, Gemini, Qwen). Auto-memory, sub-agents, agent teams, MCP. Apache 2.0.
  • CodeWhale — 2026. Terminal coding agent with 40K+ stars. 30+ providers, local models via Ollama/vLLM. TUI, headless mode, web UI. MIT license.
  • SideCar — 2026. Free, self-hosted VS Code agent extension. Full agent loop, local Ollama models, inline completions, MCP. Drop-in for Copilot/Claude Code. MIT.
  • nanobot (HKUDS) — 2026. Open-source, ultra-lightweight personal AI agent with WebUI, chat channels, MCP, memory, and scheduling. 46K★. MIT.
  • MiMoCode (Xiaomi) — Jun 2026. Terminal-native coding agent with persistent memory, subagent orchestration, and goal-driven autonomous loops. 12.4K★. MIT.

📝 Code Models

📅 Last checked: August 1, 2026

Specialized for code generation, completion, and analysis.

  • MAI-Code-1-Flash (Microsoft) — Jun 2026. Microsoft's open-weight coding model for lowering infrastructure costs.
  • DeepSeek Coder — State-of-the-art open-weight code generation. DeepSeek's coder series leads SWE-bench. MIT license.
  • Qwen2.5-Coder (Alibaba) — Highly capable code model series (1.5B–32B). Excellent balance of speed and quality. Apache 2.0.
  • Codestral (Mistral) — Mistral's dedicated code generation model — fill-in-the-middle, completion, and instruction.
  • CodeGemma (Google) — Google's Gemma architecture fine-tuned for code completion and instruction. Apache 2.0.
  • StarCoder2 (BigCode) — Transparently trained code model covering 619 languages. OpenRAIL-M license.
  • Yi-Coder (01.AI) — Efficient coding model with strong long-context understanding. Yi License (Apache 2.0 compatible).
  • Granite Code (IBM) — IBM's enterprise-grade code model, available in multiple sizes. Apache 2.0.
  • Phi-4-mini (Microsoft) — Lightweight model optimized for reasoning and code. Punches above its weight class. MIT license.
  • Qwen3-Coder-Next (Alibaba) — Early 2026. Latest generation of Qwen's code series. Strong reasoning and long-context coding capabilities. Apache 2.0.
  • CodeLlama (Meta) — Aug 2023. Llama 2-based code generation pioneer. Supports infilling, completion, and instruction. Llama 2 Community License.
  • WizardCoder (WizardLM) — 2023. Evol-Instruct fine-tuned for complex coding tasks. Strong general code generation performance. Apache 2.0.
  • OpenCodeInterpreter — 2024. Integrates execution feedback to iteratively improve generated code. Bridges generation and execution. Apache 2.0.
  • Stable Code 3B (Stability AI) — Aug 2023. Lightweight 3B code model optimized for fill-in-the-middle. Efficient for local autocompletion. StabilityAI license.
  • CodeGeeX2 (THUDM) — 2023. Multilingual code model supporting 20+ languages. Strong in both Chinese and English code tasks. Apache 2.0.
  • CodeT5+ (Salesforce) — 2023. Encoder-decoder architecture unifying code generation, completion, and understanding. BSD-3 license.
  • SantaCoder (BigCode) — 2023. Light 1.1B model specialized for Python, Java, and JavaScript. Fast and efficient for IDE integration.

🧬 Embedding Models

📅 Last checked: August 1, 2026

Free, open-weight embedding and reranker models for semantic search, RAG, and text representation.

  • Qwen3-Embedding (Alibaba) — 2K★. #1 on MTEB multilingual leaderboard. Sizes: 0.6B/4B/8B. 32K context, MRL support, instruction-aware. Includes reranker models. Apache 2.0.
  • BGE-M3 (BAAI) — Multi-lingual (100+ languages), multi-functionality (dense, sparse, colbert), multi-granularity (8K tokens). MIT.
  • FlagEmbedding (BAAI) — 12K★. Framework and model zoo: BGE series, BGE-VL (multimodal), bge-en-icl, bge-multilingual-gemma2 (9B multilingual SOTA). MIT.
  • nomic-embed-text-v2 (Nomic AI) — 1.5B MoE embedding model. 8192 context. Matches or exceeds OpenAI text-embedding-3-small. Apache 2.0.
  • mxbai-embed-large-v1 — 0.3B lightweight embedding. Top of MTEB among sub-0.5B models. Apache 2.0.

🔍 RAG & Vector Databases

📅 Last checked: August 1, 2026

Free tools for building retrieval-augmented generation pipelines — vector storage, embedding search, and document retrieval.

  • Chroma — AI-native open-source embedding database. Runs in-process, no GPU needed. GitHub
  • Qdrant — High-performance vector search engine. Free tier on Qdrant Cloud or self-host via Docker. GitHub
  • pgvector — Vector similarity search inside PostgreSQL. Free if you already run Postgres.
  • LanceDB — Developer-friendly vector database built on Lance columnar format. Runs locally, no server needed. GitHub
  • Weaviate — Open-source vector database. Free sandbox tier on Weaviate Cloud. GitHub
  • Milvus (Zilliz) — Cloud-native vector database. Free tier on Zilliz Cloud or self-host. GitHub
  • txtai — AI-powered semantic search and RAG in a single Python package. GitHub
  • R2R (SciPhi) — Production-ready RAG engine with API, user management, and observability. Note: Last release Jun 2025. Consider alternatives like Dify or LangGraph.
  • Docling (IBM) — Document understanding and conversion for RAG pipelines. Extracts PDFs, images, and more. GitHub
  • Unstructured.io — Preprocessing toolkit for documents (PDF, HTML, Word) for RAG pipelines. Free tier available.

🧩 Agentic Frameworks

📅 Last checked: August 1, 2026

Free, open-source frameworks for building AI agents and multi-agent systems.

  • LangGraph (LangChain) — Low-level framework for building stateful, multi-agent applications. GitHub
  • CrewAI — Multi-agent framework for orchestrating specialized AI agents to work together. GitHub
  • AutoGen (Microsoft) — Extensible framework for building multi-agent conversations. Note: In maintenance mode (last push Apr 2026). GitHub
  • Agno (formerly Phidata) — Full-stack AI framework for building multimodal agents with memory, knowledge, and tools. GitHub
  • PydanticAI — Agent framework by Pydantic with type-safe outputs and dependency injection. GitHub
  • Mastra — TypeScript framework for building AI applications and agent workflows. GitHub
  • OpenAI Agents SDK — Lightweight SDK for building single and multi-agent systems. GitHub
  • Semantic Kernel (Microsoft) — SDK for orchestrating AI agents with planners, memory, and connectors. GitHub
  • Dify — LLM app development platform with visual workflow builder and agent capabilities. GitHub
  • Flowise — Low-code visual LLM flow builder with drag-and-drop interface. Note: Company acquired by Workday; repo actively maintained (last push Jul 2026). GitHub
  • Fazm — Apr 2026. Open-source local computer-use agent for macOS. Drives apps via accessibility APIs, model-agnostic, faster than screenshot-based agents.
  • Smolagents (Hugging Face) — Minimalist agent library where agents "think in code." Lightweight, zero boilerplate. Supports code agents and tool-calling agents.
  • Swarms — Enterprise-grade multi-agent orchestration framework. Scalable infrastructure for autonomous agent swarms. Highly modular. Actively maintained (last push Aug 2026).
  • Letta (MemGPT) — Framework for long-term agent memory. Virtual memory management that pages data in/out of context like an OS. Persistent agents.
  • Griptape — Enterprise agent framework with strictly typed Pipelines, Workflows, and Agents. Structure-first, production-ready.
  • Atomic Agents — Framework inspired by Atomic Design. Compose agents from small, reusable, modular components. Testable and scalable.
  • PraisonAI — Low-code multi-agent framework. Define agent roles, tasks, and flows via YAML configuration. Wraps underlying agent frameworks.
  • Cognee — GraphRAG framework for agent knowledge management. Builds interconnected knowledge graphs from unstructured data.
  • MetaGPT — Multi-agent framework simulating a full software team. Assigns Agent, Product Manager, Engineer roles. Implements SOPs for end-to-end code generation. Note: Active development (last push Jan 2026).
  • ChatDev (OpenBMB) — Virtual software company driven by multi-agent collaboration. Follows waterfall model through design, coding, testing, and documentation.
  • AutoGPT — The original autonomous agent experiment. Sets its own goals, iterates on tasks, and executes without continuous human input. Web browsing and file management.
  • Bee Agent Framework (IBM) — Production-ready framework for building reliable AI agents in Python and TypeScript. Modular, with built-in observability and IBM research optimizations.
  • Eliza (elizaOS) — Multi-platform agent framework for creating character-driven AI agents. Handles social media interaction, complex decision-making, and autonomous behavior across platforms.
  • Qwen-Agent (Alibaba) — Agent framework tightly integrated with the Qwen model family. Optimized for function calling, code execution, RAG, and tool use with Qwen models.
  • AGiXT — Extensible modular AI agent automation platform. Plugin system for swapping LLMs, memory backends, and tools. Highly customizable agent workflows.
  • Microsoft Agent Framework — 12.3K★. Production-grade multi-agent framework for Python and .NET. Graph-based workflows, streaming, human-in-the-loop. MIT.
  • GenericAgent — 13.5K★. Minimal, self-evolving autonomous agent framework. ~3K lines core, 9 atomic tools. Self-crystallizing skill tree from every task. MIT.
  • Omnigent — 7.6K★. Open-source meta-harness orchestrating Claude Code, Codex, Cursor, Pi, and custom agents. Real-time collaboration from any device. Apache 2.0.

🔧 MCP Servers & Tools

📅 Last checked: August 1, 2026

Model Context Protocol (MCP) servers that connect AI assistants to external tools, data sources, and APIs.

  • GitHub MCP Server — 31.7K★. Official GitHub MCP server by GitHub. Repository management, issue/PR automation, CI/CD intelligence, code analysis. OAuth or PAT auth. MIT license.
  • GitMCP — 8.2K★. Free, open-source, remote MCP server for any GitHub project. Zero-setup documentation and code access for AI assistants. Apache 2.0.
  • MCP Reference Servers (Anthropic) — 88.9K★. Official reference implementations: Filesystem, Git, Fetch, Memory, Time, Sequential Thinking. Apache 2.0 / MIT.
  • MCP Server Toolkit — Semantic code search, docs server, database server (Postgres/MySQL/SQLite), OpenAPI introspection. One-command npx setup. MIT.
  • MCP Depot — Self-hosted MCP server hub with web UI. Connect Jira, GitHub, Confluence, Jenkins, custom APIs. AGPL-3.0.
  • free-search-mcp — Free web search via MCP with no API key required. Multiple search backends with automatic failover. Apache 2.0.

🎛 Fine-tuning Tools

📅 Last checked: August 1, 2026

Tools to fine-tune free models on your own data — all free and open-source.

  • Unsloth — Fast memory-efficient fine-tuning. 2x faster, 50% less memory. Supports QLoRA, LoRA, full fine-tune.
  • Axolotl — Streamlined fine-tuning framework supporting multiple model architectures and quantization methods.
  • LLaMA-Factory — Easy-to-use fine-tuning with web UI. Supports 100+ models, multiple training methods.
  • Hugging Face TRL — Transformer Reinforcement Learning library. SFT, PPO, DPOTrainer, GRPOTrainer for aligning models.
  • XTuner (InternLM) — Efficient fine-tuning toolkit supporting QLoRA, LoRA, and full fine-tune with multiple model architectures.
  • Ludwig (Predibase) — Declarative ML framework. Fine-tune models with a simple config file. GitHub

✨ Prompt Engineering Tools

📅 Last checked: August 1, 2026

Free tools for testing, managing, and optimizing prompts.

  • Promptfoo — Company acquired by OpenAI; repo actively maintained (last push Aug 2026). Open-source tool for prompt testing and evaluation. Systematic A/B testing of prompts. GitHub
  • Fabric (Daniel Miessler) — Open-source framework for augmenting humans with AI. Library of curated prompts (patterns) for common tasks.
  • LangFuse — Open-source LLM engineering platform with prompt management, versioning, and evaluation. GitHub
  • DSPy (Stanford) — Framework for algorithmically optimizing LM prompts and weights. GitHub
  • Agenta — Open-source LLM platform for prompt management, evaluation, and deployment. GitHub

📊 LLM Evaluation & Observability

📅 Last checked: August 1, 2026

Free, open-source tools for tracing, evaluating, and monitoring LLM applications in development and production.

  • Langfuse — 31.5K★. Full LLM engineering platform: tracing, evaluations, prompt management, playground, datasets. Self-hostable. MIT (core). GitHub
  • Opik (Comet) — 20.7K★. Open-source LLM observability, evaluation, and agent tracing. Datasets, experiments, LLM-as-judge, guardrails, prompt management. Apache 2.0.
  • Phoenix (Arize AI) — AI observability platform with OpenTelemetry-based tracing, evals, experiments, and prompt playground. Elastic License 2.0.
  • TruLens — 3.4K★. Agent-specific evaluations (7 purpose-built evaluators). OpenTelemetry tracing, MCP support, batch and inline evaluation. MIT.
  • OpenLLMetry (Traceloop) — 7K★. OpenTelemetry-based LLM observability. Send traces to any OTLP-compatible backend. Apache 2.0.

📊 Datasets

📅 Last checked: August 1, 2026

Free, open datasets for training, fine-tuning, and evaluating models.

☁ Model Hosting Platforms

📅 Last checked: August 1, 2026

Free platforms that host models — run inference without downloading anything.

📚 Learning Resources

📅 Last checked: August 1, 2026

Free courses, books, and tutorials for learning AI and LLMs.

🏆 Resources & Leaderboards

📅 Last checked: August 1, 2026

👥 Communities

📅 Last checked: August 1, 2026

License

CC0

To the extent possible under law, the author has waived all copyright and related or neighboring rights to this work.

About

A curated list of free AI models, APIs, and tools you can use without paying a cent.

Resources

Readme
License

Contributing

Contributing
Activity

Stars

1.5k stars

Watchers

21 watching

Forks

134 forks
Report repository

Releases

Packages

Contributors