Dify
Last reviewed
Jun 4, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,307 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,307 words
Add missing citations, update stale details, or suggest a clearer explanation.
Dify (stylized Dify.AI) is an open-source platform for building, deploying, and operating applications and agents powered by large language models. Developed by LangGenius, Inc., it combines a visual workflow builder, a prompt IDE, a retrieval-augmented generation pipeline, an AI agent framework, model management, and observability into a single self-hostable system, a category often described as LLMOps. The project was open-sourced on GitHub on 15 May 2023 and grew into one of the most-starred repositories on the platform, passing 100,000 stars in June 2025 and exceeding 144,000 by mid-2026. Founded by Luyu Zhang and John Wang, LangGenius raised a 30 million US dollar Series Pre-A round in March 2026 at a reported 180 million US dollar valuation. The name "Dify" is a contraction the company glosses as "Do It For You" (and elsewhere as a blend of "Define" and "Modify").
Dify positions itself as a "production-ready platform for agentic workflow development," aiming to take teams from prototype to production without forcing them to assemble their own stack of orchestration code, vector stores, prompt-management tooling, and serving infrastructure. According to the project README, "Dify is an open-source LLM app development platform. Its intuitive interface combines AI workflow, RAG pipeline, agent capabilities, model management, observability features and more, letting you quickly go from prototype to production."
The platform is model-agnostic: it integrates hundreds of proprietary and open-source models from dozens of inference providers, including OpenAI's GPT family, Anthropic's Claude, Google Gemini, Mistral, Meta's Llama, Azure OpenAI, Hugging Face, Replicate, and any OpenAI-API-compatible endpoint, as well as locally hosted models served through tools such as Ollama. Applications built on Dify expose APIs so they can be embedded into existing products, a backend-as-a-service model the company emphasizes for enterprise integration.
LangGenius was founded in 2023 by Luyu Zhang (founder and chief executive) and John Wang (co-founder and chief engineer, known on GitHub as "takatost"). Before Dify, Zhang held product leadership roles in China, most notably as product director of CODING, the DevOps platform owned by Tencent; earlier he had founded Feie Work, a testing-process management and collaboration tool. He began experimenting with generative AI in 2022 and started Dify the following year with the stated belief that the technology "should reach everyone equally."
The team open-sourced Dify on 15 May 2023, shortly after the wave of interest that followed the launch of ChatGPT. The product was designed from the start as a self-hostable platform rather than a closed SaaS, which contributed to rapid adoption among developers who wanted to keep model usage and data inside their own infrastructure. The company is incorporated in the United States and operates from the San Francisco Bay Area, with Zhang relocating to the US as the company expanded internationally; engineering and a substantial part of the community remain global.
Dify's GitHub repository (langgenius/dify) became one of the platform's fastest-growing projects. The company announced reaching 100,000 stars on 5 June 2025, describing the project as among the top 100 open-source repositories worldwide; by mid-2026 the repository had roughly 144,000 stars and more than 22,000 forks, with over 800 contributors. The company has reported the software running on more than 1.4 million machines across over 175 countries and territories, with more than 5 million downloads.
Notable timeline events include the release of Dify v1.0.0 on 28 February 2025, which introduced a plugin system that decoupled models and tools from the core platform and launched the Dify Marketplace. The company was named AWS Social Impact Partner of the Year in 2025 and was featured at NVIDIA GTC 2025.
LangGenius raised an undisclosed seed round in mid-2023. Its first widely publicized institutional financing came in early 2026.
| Round | Date | Amount | Valuation | Lead / notable investors |
|---|---|---|---|---|
| Seed | June 2023 | Undisclosed | Not disclosed | Delian Capital, FutureX Capital (per Crunchbase) |
| Series Pre-A | 9 March 2026 | 30 million USD | ~180 million USD (reported) | Led by HSG (Hongshan); with GL Ventures, Alt-Alpha Capital, 5Y Capital, Mizuho Leaguer Investment, NYX Ventures |
The Series Pre-A round was led by HSG (Hongshan, the firm formerly known as Sequoia Capital China), with participation from GL Ventures (the early-stage arm of Hillhouse Investment), Alt-Alpha Capital (described in announcements as a spin-out from Bessemer Venture Partners), 5Y Capital, Mizuho Leaguer Investment, and NYX Ventures. Crunchbase has tallied total funding for the company at roughly 41.5 million US dollars across its rounds. In announcing the round, Zhang framed the market shift the company is targeting: "AI is moving from demos to production systems, and teams need infrastructure they can operate with confidence."
Dify bundles several components that would otherwise be separate tools. The core feature set, as documented by the project, includes the following.
| Component | What it does |
|---|---|
| Workflow | A visual canvas for building and testing multi-step AI pipelines, branching logic, and orchestration without writing the glue code by hand. |
| Comprehensive model support | Integration with hundreds of proprietary and open-source models across dozens of inference providers, plus self-hosted and OpenAI-API-compatible endpoints. |
| Prompt IDE | An interface for authoring prompts, comparing outputs across models side by side, and adding capabilities such as text-to-speech to chat apps. |
| RAG pipeline | End-to-end document ingestion, chunking, embedding, indexing, and retrieval, with out-of-the-box text extraction from PDFs, PowerPoint, and other formats; supports vector, keyword/BM25, and hybrid retrieval with optional reranking. |
| Agent capabilities | Agents built on LLM function calling or the ReAct pattern, with more than 50 built-in tools and the ability to add custom ones. |
| LLMOps / observability | Monitoring and analysis of application logs and performance over time; integrations with observability tools including Langfuse, Opik, and Arize Phoenix. |
| Backend-as-a-Service | Every application exposes APIs so it can be embedded into business logic and external products. |
Dify lets builders create several kinds of applications from the same console: simple chat assistants, text-completion apps, "agent" apps that can call tools autonomously, and Workflow / Chatflow apps that orchestrate complex multi-node logic on the visual canvas. The Workflow canvas chains nodes such as LLM calls, knowledge retrieval, code execution, HTTP requests, conditional branches, and tool invocations.
The knowledge base is one of Dify's strongest differentiators. Users upload documents (PDF, Word, Notion pages, web URLs, and more), and Dify handles chunking, embedding, and indexing automatically, letting the user choose the embedding model and vector store while the platform manages the pipeline. In 2025 the company introduced a Knowledge Pipeline that brings the RAG extract-transform-load path onto the same visual canvas, exposing each step (parsing, cleaning, chunking, enrichment, vectorization) as a configurable node and connecting to sources such as Google Drive, Notion, Confluence, AWS S3, and Slack through marketplace plugins.
With v1.0.0 (February 2025), Dify shipped a plugin architecture that decouples models, tools, and agent strategies from the core, distributed through the Dify Marketplace, which launched with more than 120 plugins. Plugin types include models (for example OpenAI o1-series, Gemini 2.0-series, and DeepSeek-R1), tools (Perplexity, Slack, Firecrawl, Jina AI, Stability, ComfyUI, and others), agent strategies, and extensions. Dify also added support for the Model Context Protocol (MCP), allowing it to call external HTTP-based MCP services and to expose its own apps as MCP endpoints.
Dify is delivered as a set of services (an API/backend, a worker, a web frontend, a database, Redis, and a vector store) typically run via Docker Compose. In 2025 the company rebuilt its execution layer around a queue-based graph engine, in which workflow nodes enter a unified scheduler that manages dependencies and parallel execution to reduce errors in concurrent tasks. The minimum recommended footprint for self-hosting is modest, with the project listing requirements of at least 2 CPU cores and 4 GiB of RAM.
Dify is frequently compared to developer frameworks such as LangChain and to visual builders such as Flowise and Langflow.
Dify can be used in several ways.
| Edition | Description |
|---|---|
| Dify Cloud | A fully hosted SaaS at cloud.dify.ai with a free sandbox tier and paid plans. |
| Community Edition | The open-source self-hosted version, deployed via Docker Compose (and community Kubernetes/Helm charts). |
| Premium / Enterprise | Paid self-hosted offerings adding multi-tenant management, single sign-on (SAML, OIDC, OAuth2), model load balancing, and support; Enterprise is built for Kubernetes with an official Helm chart for private-cloud or on-premises deployment to meet compliance and data-residency requirements. |
Dify Enterprise is also distributed through the Microsoft Azure Marketplace and AWS Marketplace. Enterprise inquiries are handled through the company's sales channel.
Dify Cloud's published plans (as of 2026) are tiered by message credits, applications, knowledge documents, storage, and team seats.
| Plan | Price | Highlights |
|---|---|---|
| Sandbox | Free | 200 message credits, 1 workspace, 1 member, 5 apps, 50 knowledge documents, 50 MB knowledge storage. |
| Professional | ~59 USD per workspace / month | 5,000 monthly message credits, 3 members, 50 apps, 500 knowledge documents, 5 GB storage, no API rate limits. |
| Team | ~159 USD per workspace / month | 10,000 monthly message credits, 50 members, 200 apps, 1,000 knowledge documents, 20 GB storage. |
Annual billing is offered at a discount, and self-hosted/enterprise pricing is quoted on request. Because Dify is model-agnostic, users typically bring their own model API keys and pay model providers directly for inference; "message credits" govern certain hosted features rather than reselling raw tokens.
The company reports that more than 2,000 teams and over 280 enterprises run on commercial versions of Dify. Named adopters cited in its 2026 funding announcement and marketing include Maersk, ETS (Educational Testing Service), Anker Innovations, Novartis, Volvo Cars, Thermo Fisher Scientific, and Ricoh. Reported use cases span document-review pipelines, internal copilots grounded in enterprise knowledge, customer-support automation with escalation paths, and operational workflows such as invoice auditing and correspondence drafting. The company has described a customer Q&A deployment serving more than 19,000 employees across over 20 departments. These figures are largely self-reported by the company and its customers.
Dify's source code is released under the Dify Open Source License, which is the Apache License 2.0 with two additional conditions:
web/ directory or the "web" Docker image), the LOGO and copyright information in the console and applications may not be removed or altered. This condition does not apply to uses that do not involve the frontend.All other rights and obligations follow the standard Apache 2.0 terms. The contributor terms note that contributors permit Dify to adjust licensing and to use contributed code commercially, including in its cloud service. The license is therefore best described as "open source with commercial-use conditions" rather than a pure permissive license; commentators have debated whether it meets a strict definition of free and open-source software because of the added restrictions.
Large language model - Retrieval-augmented generation - AI agents - Agentic AI - LangChain - Prompt engineering - Open source - Generative AI