Agenta is an open-source workspace for building, running and automating AI agents (sometimes called AI coworkers). Teams can build agents through chat, give them skills and tool connections via MCP servers, and deploy them in Slack, WhatsApp or Telegram. It supports self-hosted and cloud deployment, lets agents use their own browser and files, and includes approvals and permissions so people stay in control of what agents do. It is aimed at teams across engineering, sales, support and operations who want to automate repetitive research, drafting, review or reporting tasks while keeping a human in the loop for review.
Best Dify Alternatives in 2026Free, Open-Source & Cheaper Options
Dify is a local llms & self-hosting tool — Dify is a platform for building production-ready AI agents, agentic workflows, and knowledge pipelines on a single visual canvas. It lets teams design agents, connect models and tools, and build searchable knowledge bases without starting from scratch in code. Dify can be used as a managed cloud service, deployed in a private VPC, or self-hosted via its open-source Community Edition. It suits independent developers, small teams, and larger enterprises that need governance, SSO, and audit controls alongside the ability to build and ship AI applications quickly. Below are the highest-rated alternatives in the same space, ranked by community rating, with pricing transparency so you can find a cheaper or fully free swap.
Dify alternatives
Mistral AI is a European AI company offering open-weight large language models alongside a platform for enterprises and individuals to build, customize, and deploy AI systems. Its products include Vibe, a personal AI assistant for chat, coding, and everyday tasks; Studio, a platform to build and govern AI agents; Forge, for model customization and fine-tuning; and Inference and Compute offerings for running models on Mistral's cloud, in a private VPC, or on-premises. It is aimed at enterprises, developers, and individuals who want control over their AI models and data, with customers in energy, manufacturing, financial services, and public sector industries.
vLLM is an open-source inference and serving engine for large language models, built for high throughput and efficient memory use. It is aimed at developers and organizations who need to deploy open-source LLMs in production, whether on GPUs, NPUs, TPUs, or CPUs across multiple hardware vendors. It provides a drop-in OpenAI-compatible API so existing applications can integrate with minimal changes. The project focuses on maximizing hardware utilization to reduce inference costs while supporting a wide and growing range of trending open-source models from providers like DeepSeek, Meta, Mistral AI, Qwen, and others.
LLM is an open source command-line tool and Python library for working with large language models, both remote APIs and models that run locally on your own machine. It supports providers such as OpenAI, Anthropic's Claude, Google's Gemini, and many others through a plugin system, as well as locally installed models via plugins like llm-ollama. It is aimed at developers and technical users who want to run prompts, build pipelines, extract structured data, or manage embeddings from the terminal or from Python scripts, without being locked into a single provider.
RWKV 5 is an open source AI language model project built around the RWKV architecture, which combines ideas from recurrent neural networks and transformers. It offers linear time processing, constant memory use regardless of context length, and no key-value cache, aiming to deliver transformer-level performance with RNN efficiency. The project is attention-free and is part of the Linux Foundation AI projects. It provides model weights, training code, a finetuning toolkit, runners, and apps for multiple platforms. It is intended for developers, researchers, and hobbyists who want to train, finetune, or run large language models locally or with efficient inference, rather than for non-technical end users looking for a polished product.
dlite-v2-1_5b is an open source language model hosted on Hugging Face, created by AI Squared. It is derived from OpenAI's GPT-2 large model and fine-tuned on Databricks' databricks-dolly-15k dataset to give it chat-style, instruction-following behavior. It is licensed under Apache v2.0 for both research and commercial use. The model is aimed at developers, researchers, and hobbyists who want to experiment with a small, cheaply trained language model that still shows interactive, conversational capabilities. It is not a state-of-the-art model, but is meant to demonstrate that useful AI capability can be built on a small scale. Users can download it directly or run it through standard model-serving tools such as Transformers, vLLM, SGLang, or Docker.
ChatGLM3-6B is an open-source bilingual (Chinese and English) large language model with 6 billion parameters, hosted on Hugging Face by Z.ai (THUDM). It is built on the GLM architecture and supports multi-turn dialogue, function calling, code interpretation, and agent-style tasks. The model is distributed as weights that can be downloaded and run locally using the Transformers library, making it suited for developers, researchers, and organizations who want a self-hosted conversational AI model rather than a hosted API product. Alongside the chat model, a base model and a long-context 32K variant are also released. It is free for academic research and, after registration, free for commercial use.
Jais-13b-chat is a 13 billion parameter bilingual large language model for Arabic and English, hosted on Hugging Face by Inception42. It is built on a GPT-3 style transformer decoder architecture with SwiGLU activations and ALiBi position embeddings for handling long sequences. The base model, Jais-13b, was pretrained on 116 billion Arabic tokens and 279 billion English tokens, then fine-tuned on a curated set of 4 million Arabic and 6 million English prompt-response pairs, with additional safety-oriented instruction tuning. It is intended for developers and researchers who want an open, self-hostable chat model with a particular focus on Arabic language and Arab world topics, alongside English support.
CC BY-SA 4.0 is a public copyright license provided by Creative Commons that allows anyone to copy, redistribute, remix, transform, and build upon a creator's work, even for commercial purposes. It is not a software tool but a legal license meant for creators who want to share their work openly while requiring that any adaptations be distributed under the same license. It is relevant to people publishing text, images, code, data, or other creative works who want others to be able to reuse and build on their material under clear, standardized terms, including attribution and share-alike requirements.
Cerebras is the organization behind the Wafer-Scale Engine, a custom processor used in its Cerebras CS-X systems. On Hugging Face, Cerebras maintains a profile where it publishes open models it has trained and supports models that use Cerebras as an inference provider. The Hugging Face page lets users browse popular models enabled by Cerebras, explore models trained on Cerebras hardware, and join a Discord community to discuss the company's work and research. This listing is aimed at machine learning engineers, researchers, and developers who want to use or study large language models trained with Cerebras hardware, or who want to run inference through Cerebras-backed models rather than at the hardware company's own product pages.
Zephyr 7B (zephyr-7b-gemma-v0.1) is a text generation model hosted on Hugging Face, built by fine-tuning a Gemma base model using direct preference optimization (DPO) on the argilla/dpo-mix-7k dataset. It is designed for conversational and general text generation tasks. The model is released under a gemma-terms-of-use license and can be downloaded, explored, and deployed directly from its Hugging Face model page. It is aimed at developers, researchers, and ML practitioners who want to run or fine-tune an open large language model, either locally or through supported inference providers, rather than at non-technical end users looking for a ready-made chat app.
Phi-2 is an open-source text generation model with 2.7 billion parameters, built by Microsoft and hosted on Hugging Face. It is a Transformer trained on the same data as Phi-1.5 plus additional synthetic NLP texts and filtered websites chosen for safety and educational value. It is not fine-tuned with reinforcement learning from human feedback. It is best suited for prompts in QA, chat, or code format. The model is aimed at researchers and developers who want a small, unrestricted model to study topics like reducing toxicity, understanding bias, and improving controllability, and who want to run or fine-tune a compact LLM locally or via common inference tooling rather than relying on a large hosted API.
Nous Research is an open-source AI research organization focused on training language models and building AI agent tools. Its main product, Hermes Agent, is described as a widely used open source agent harness with a built-in learning loop that creates and improves skills from experience and builds a model of the user across sessions. Hermes is available through a terminal interface, a desktop application for Mac, Windows and Linux, and a cloud portal that can host personal agents running continuously in their own container. It is aimed at developers and users who want to run or build self-improving AI agents, with options to use them locally or in the cloud, and connect across devices while keeping the same chat history and settings.
Jan is an open-source desktop application that works as an alternative to ChatGPT, letting people run AI models on their own computer instead of relying entirely on cloud services. Users can download open models such as Llama, Mistral, Qwen, Gemma, DeepSeek, or Kimi and run them locally and offline, or connect to cloud model providers like OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini. It is built for developers, privacy-conscious users, and AI enthusiasts who want control over which models they use and where their data goes. Jan is free, open source, and available for download with millions of downloads to date.
Llama-3-70B-Instruct is a 70-billion-parameter text generation model from Meta, fine-tuned for instruction following and conversational use. It is hosted on Hugging Face, where it can be downloaded or run through supported inference providers. The model is built on the Llama 3 architecture, uses the Transformers library with PyTorch and Safetensors weights, and is licensed under the Llama 3 license with gated access requiring profile sharing. It is intended for developers and researchers who want to run or fine-tune a large open-weight language model for tasks like chat, coding assistance, and general text generation, either locally or through compatible serving frameworks such as vLLM and SGLang.
LM Studio is a desktop application for downloading and running large language models (LLMs) locally on your own computer, using the LM Studio runtime with llama.cpp and MLX. It includes Bionic, an agent built for work and coding tasks such as creating and editing documents, slides, PDFs, and software. Bionic also offers real-time voice transcription processed locally. Users can run local models for free, or optionally connect to cloud-hosted open and frontier models for more demanding tasks. It is aimed at developers, privacy-conscious users, and anyone who wants to run AI models on their own hardware instead of relying solely on cloud services.
GPT-NeoX-20B is a 20 billion parameter autoregressive language model developed by EleutherAI, trained on the Pile dataset using the GPT-NeoX library. Its architecture closely resembles GPT-3 and is nearly identical to GPT-J-6B, using Rotary Position Embedding and a 2048 token sequence length. The model is open source under the Apache 2.0 license and hosted on Hugging Face, where it can be downloaded and run through libraries such as Transformers, vLLM, SGLang, or Docker. It was developed primarily for research purposes, aimed at researchers and developers who want to study or build on an open autoregressive language model, extract internal representations, or use it as a base for downstream text generation tasks.
fastchat-t5-3b-v1.0 is an open-source chatbot model hosted on Hugging Face, built by the FastChat developers (Dacheng Li, Lianmin Zheng, Hao Zhang). It is created by fine-tuning Flan-T5-XL, a 3 billion parameter encoder-decoder transformer, on 70K conversations collected from ShareGPT. The model can autoregressively generate responses to user inputs and is released under the Apache License 2.0. It is intended for entrepreneurs and researchers in natural language processing, machine learning, and artificial intelligence who want to build commercial chatbots or conduct NLP research, and it can be loaded directly through the Hugging Face Transformers library.
Falcon-180B is a large open text generation model released by the Technology Innovation Institute (TII) and hosted on Hugging Face. It has 180 billion parameters and was trained on the RefinedWeb dataset, with support for English, German, Spanish and French. It is built on the Transformers library and distributed as safetensors files. The model is aimed at developers and researchers who want to run or fine-tune a large language model locally or on their own infrastructure, using tools such as Transformers, vLLM, SGLang, or Docker. Access requires agreeing to the Falcon-180B TII license and acceptable use policy before downloading.
OpenHermes-13B is an open source large language model fine-tuned by teknium on the Hermes dataset, hosted on Hugging Face. It is built on Llama-2-13B and trained on 242,000 entries of primarily GPT-4 generated instruction and conversation data, drawn from a mix of open datasets including GPTeacher, WizardLM, Airoboros, Camel-AI, CodeAlpaca, and GPT4-LLM. The model is intended for developers and researchers who want a self-hostable text generation model for instruction following, roleplay, and code assistance tasks. It is released under the MIT license and can be downloaded and run locally or deployed via common inference tooling such as Transformers, vLLM, SGLang, and Docker.
RedPajama-INCITE is a family of open-source large language models published on Hugging Face by Together, an organization building decentralized computing infrastructure for foundation models. The models are hosted under the togethercomputer organization page, alongside other open models and datasets such as HybridAgent, Tev1, and StripedHyena. RedPajama-INCITE is intended for developers and researchers who want openly available base and instruction-tuned language models they can download, run locally, fine-tune, or self-host rather than relying on a closed commercial API. It fits into the Local LLMs & Self-Hosting category of tools because it can be deployed outside of any single vendor's hosted service.
RunThisLLM is a lookup tool for people who want to know whether a given local AI model will run on their computer, or what hardware they would need to run a specific model. It covers more than 295 local LLMs as well as image, video, and audio AI models, offering build specs, performance estimates, and benchmarks for each. Users can search by the hardware they already own to see which models it can handle, or search by a model they want to run to see what hardware is required. It also offers a custom build request service and a Chrome extension for quick lookups. It's aimed at hobbyists, developers, researchers, and anyone considering a local AI setup.
OpenLM 7B (open_lm_7B_1.25T) is a machine learning model hosted on Hugging Face, released by the ML Foundations organization under the MIT license. It is a 7-billion-parameter language model, with the name indicating it was trained on roughly 1.25 trillion tokens. The model page provides file downloads and version history, but the README is empty, so no further detail about training data, architecture, or intended use is provided by the publisher. It is suited for developers, researchers, and ML practitioners who want to download and run an open-source large language model themselves rather than use a hosted API, as the model does not currently have inference provider support on Hugging Face.
OpenAssistant Pythia (oasst-sft-4-pythia-12b-epoch-3.5) is an open-source text-generation model hosted on Hugging Face. It is the fourth supervised-fine-tuning iteration of the Open-Assistant project, built on EleutherAI's Pythia-12B-deduped base model and fine-tuned on human-collected assistant conversation demonstrations gathered through the Open-Assistant web app before March 25, 2023. It is intended for developers, researchers and hobbyists who want a self-hostable, instruction-following conversational language model they can run locally or deploy themselves, rather than relying on a closed commercial chat API. The model is distributed under the Apache 2.0 license and can be downloaded and run through common machine learning tooling.