Open-source and local AI alternatives
You can run capable AI on your own computer or server. This page separates what is truly open source from what is merely free or downloadable, and shows honest trade-offs.
Last reviewed . AI products change quickly; check the linked sources before relying on details.
Five different things people call open
| Term | Meaning |
|---|---|
| Open-source software | Code under a licence approved as open source (for example MIT, Apache-2.0). You can read, change and share it. Examples here: Ollama, llama.cpp, Letta. |
| Open-weight models | Trained model files you can download and run. The licence may restrict use, and training data and code are often not published, so many are not open source in the OSI sense. |
| Self-hosted | Software or models you run on your own computer or server, whatever the licence. You control the data and own the maintenance. |
| Open APIs | Documented interfaces anyone can call, sometimes with free tiers. The service itself may be closed. |
| Free hosted | A provider runs it and does not charge, within limits. Convenient, but terms, limits and data use are the provider's. |
The Open Source Initiative maintains the definition of open source. Licences with field-of-use limits, branding clauses or usage thresholds do not meet it, even if the code or weights are free to download. This page labels such cases (for example Open WebUI's branding clause, n8n's fair-code licence, and LM Studio, which is proprietary).
Production warning
Read the licence before commercial use. Licence statements below were read on 2026-10-17 for software. For model families, licences vary by release, so check the model card each time.
A. Alternatives to personal assistants
A private chat assistant needs two parts: a way to run a model and an interface. Ollama or LM Studio provides the first; Open WebUI provides a browser interface on top. This replaces plain chat well. It does not replace the integrations (calendar, email, mobile) of commercial assistants without extra work.
| Product | Replaces or overlaps | Deployment | Licence | Hardware | API dependencies | Setup | Privacy | Costs | Limitations | Best for |
|---|---|---|---|---|---|---|---|---|---|---|
| Ollama Ollama. Verified 2026-10-17 | Hosted model APIs. Runs open-weight models locally with a simple command line and local API. | Local on macOS, Windows, Linux | MIT | Depends on the model; 8 to 16 GB memory handles small models. | None for local models | Beginner. Estimate: installer plus one command to download and run a model. | Prompts stay on your machine for local models. Check any optional cloud features separately. | Free software; hardware and electricity. | Quality capped by the model you can fit; not a full chat product by itself. | First local setup. |
| LM Studio (proprietary) LM Studio. Verified 2026-10-17 | Desktop chat apps. Desktop app to download and chat with local models. | Local desktop app | Proprietary, free to download. Not open source. | Depends on model. | None for local models | Beginner. Estimate: graphical installer and model browser. | Local inference; read the app's own privacy and terms pages. | Check the vendor's current terms. | Closed source, so you rely on the vendor. | Non-technical users who want a local chat app. |
| Open WebUI Open WebUI. Verified 2026-10-17 | ChatGPT-style web interface. Self-hosted chat front end for Ollama and OpenAI-compatible APIs, with documents and users. | Self-hosted (often Docker) | Custom licence with a branding clause. Not OSI open source. Read it before commercial use. | Light for the interface; models need their own resources. | A model backend such as Ollama or an API | Intermediate. Estimate: Docker and a configured model backend. | Data stays on your server; you manage access and backups. | Free software; hosting costs. | You maintain updates and security. Licence terms limit rebranding. | Teams wanting a private chat front end. |
B. Alternatives to persistent agents
A framework is not a ready application. Letta gives you building blocks for agents with memory; you still write the agent, choose the tools and add approval steps. n8n suits fixed workflows with an AI step, which are often safer than an open-ended agent. browser-use lets a model operate a browser, and needs care around site terms and actions.
Caution: Flowise was archived in August 2026 and is not recommended. Check a project's maintenance status before depending on it.
| Product | Replaces or overlaps | Deployment | Licence | Hardware | API dependencies | Setup | Privacy | Costs | Limitations | Best for |
|---|---|---|---|---|---|---|---|---|---|---|
| Letta Letta. Verified 2026-10-17 | Persistent assistants with memory. Framework for agents with long-term memory and state. A framework, not a ready-made personal assistant. | Self-hosted or Letta's hosted service | Apache-2.0 | Server; models via API or local. | A model provider | Advanced. Estimate: requires development to build a working agent. | Self-hosting keeps memory under your control; hosted option has its own terms. | Free software; model and hosting costs. | You design the agent, tools and safeguards. | Developers prototyping agents with memory. |
| n8n (fair-code) n8n. Verified 2026-10-17 | Zapier, Make. Visual workflow automation with AI steps and many integrations. | Self-hosted or n8n cloud | Sustainable Use License (fair-code). Not OSI open source; check limits on commercial use. | Small server. | Accounts and APIs for connected services; a model provider for AI steps | Intermediate. Estimate: Docker or cloud signup, then building workflows visually. | Self-hosted keeps workflow data on your server. | Free to self-host within the licence; cloud is paid. | Licence restrictions; you maintain security. | Fixed workflows with an AI step. |
| browser-use browser-use. Verified 2026-10-17 | Browser-operating agents. Lets an AI model control a web browser to complete tasks. | Local Python library | MIT | Ordinary computer; model via API or local. | A model provider | Advanced. Estimate: Python programming required. | Page content is sent to the model you choose. | Free software; model usage costs. | Fragile on changing sites; respect site terms; add human approval for actions. | Developers automating web tasks they are allowed to automate. |
Read the agents guide for approvals and failure handling.
C. Alternatives to commercial models
Open-weight families can handle many everyday tasks. Do not choose from a single benchmark: test on your own work. Three practical limits apply.
- Hardware: larger models need much more memory. Quantisation (storing weights with fewer bits) lets bigger models fit, at some cost in quality.
- Gap: the best closed models are often ahead on hard tasks; the gap depends on the task.
- Licence: open weights does not always mean unrestricted commercial use.
See local AI and API vs local for sizing and cost.
| Product | Replaces or overlaps | Deployment | Licence | Hardware | API dependencies | Setup | Privacy | Costs | Limitations | Best for |
|---|---|---|---|---|---|---|---|---|---|---|
| llama.cpp ggml-org community. Verified 2026-10-17 | Hosted inference. Efficient inference engine for quantised models on CPUs and GPUs. | Local or self-hosted server | MIT | Runs on CPU; GPU speeds it up. | None | Advanced. Estimate: command-line build or binaries and manual model files. | Local by default. | Free software. | Lower-level than Ollama; more configuration. | Developers who want control and performance tuning. |
| Llama family Meta. Verified 2026-10-17 | Mid-size commercial models. General chat, summarisation and coding at several sizes. Released under the Llama community licence, which is not OSI open source. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| Qwen family Alibaba Cloud. Verified 2026-10-17 | Mid-size commercial models. Broad range of sizes, including coding and multilingual models. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| DeepSeek family DeepSeek. Verified 2026-10-17 | Commercial reasoning models. Reasoning and coding models; full-size versions need serious hardware. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| Mistral family Mistral AI. Verified 2026-10-17 | Small commercial models. Efficient models; some releases are open-weight and some are commercial only. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| Gemma family Google. Verified 2026-10-17 | Small commercial models. Small models designed to run on modest hardware. Gemma is separate from Gemini. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| GLM family Z.AI. Verified 2026-10-17 | Commercial coding and agent models. General and coding models with open weights. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
| Kimi family Moonshot AI. Verified 2026-10-17 | Commercial long-context models. Large models; running them locally typically needs substantial hardware or a hosted provider. | Local, self-hosted server, or hosted by third parties | Check the model card licence for each release; terms differ by version and may restrict commercial use or large deployments. | Depends on model size and quantisation; smaller sizes run on a recent laptop, larger ones need a GPU or a hosted API. | None when run locally | Intermediate. Estimate: needs a runtime such as Ollama or llama.cpp and a choice of size and quantisation. | Weights run on your hardware, so prompts stay local. Third-party hosts have their own terms. | No licence fee for most releases; hardware or hosting costs apply. | Open weights are not the same as open-source software; training data and code are often not released. Do not rank on a single benchmark. | People who need control over data, cost or customisation. |
E. Creative and multimodal
Images, video and speech can run locally or through APIs. Local gives privacy and no per-use fee, but needs a capable GPU for images and video. APIs are easier and often higher quality, with usage costs and provider terms. Model weights for each image or video model carry their own licences, separate from the software that runs them.
| Product | Replaces or overlaps | Deployment | Licence | Hardware | API dependencies | Setup | Privacy | Costs | Limitations | Best for |
|---|---|---|---|---|---|---|---|---|---|---|
| ComfyUI Comfy-Org. Verified 2026-10-17 | Hosted image and video generators. Node-based interface for local image and video generation. | Local | GPL-3.0 | A capable GPU is strongly recommended. | None for local models | Advanced. Estimate: model downloads and node graphs take time to learn. | Local generation. | Free software; hardware. | Each image or video model has its own licence; software GPL does not cover the weights. | Creators wanting control over image pipelines. |
| Whisper OpenAI. Verified 2026-10-17 | Hosted transcription services. Speech-to-text, runs locally. | Local | MIT | CPU works for small models; GPU is faster. | None | Intermediate. Estimate: Python install, or use a packaged app built on it. | Audio stays local. | Free software. | Accuracy varies by language, accent and audio quality; review transcripts. | Private transcription. |
| Piper (successor project) Open Home Foundation (OHF-Voice). Verified 2026-10-17 | Hosted text-to-speech. Fast local text-to-speech. The original repository is archived; the successor is piper1-gpl. | Local | GPL-3.0 (successor); voices have their own licences. | Runs on modest hardware. | None | Intermediate. Estimate: command line and voice downloads. | Local. | Free. | Voice quality below top hosted voices; check each voice licence. | Offline text-to-speech. |
| Kokoro hexgrad. Verified 2026-10-17 | Hosted text-to-speech. Small open text-to-speech model. | Local | Apache-2.0 | Light. | None | Intermediate. Estimate: Python install. | Local. | Free. | Limited voices and languages compared with hosted services. | Lightweight local narration. |
F. Choosing a path
- Want a private chat on a laptop: Ollama or LM Studio, then Open WebUI if you want a browser interface.
- Want automation with an AI step: n8n, within its licence.
- Want a custom agent: a framework such as Letta, and accept that you build and own it.
- Want the best quality with little effort: a hosted app, using business terms for sensitive data.
Answer a few questions in find the right AI.
Sources
- Ollama repository or site (licence read via GitHub API) (checked 2026-10-17)
- llama.cpp repository or site (licence read via GitHub API) (checked 2026-10-17)
- LM Studio (proprietary) repository or site (licence read via GitHub API) (checked 2026-10-17)
- Open WebUI repository or site (licence read via GitHub API) (checked 2026-10-17)
- Letta repository or site (licence read via GitHub API) (checked 2026-10-17)
- browser-use repository or site (licence read via GitHub API) (checked 2026-10-17)
- n8n (fair-code) repository or site (licence read via GitHub API) (checked 2026-10-17)
- ComfyUI repository or site (licence read via GitHub API) (checked 2026-10-17)
- Whisper repository or site (licence read via GitHub API) (checked 2026-10-17)
- Piper (successor project) repository or site (licence read via GitHub API) (checked 2026-10-17)
- Kokoro repository or site (licence read via GitHub API) (checked 2026-10-17)
- Open Source Initiative: Open Source Definition (checked 2026-10-17)
Frequently asked questions
What is the difference between open source and open-weight AI?
Open-source software uses a licence approved as open source, such as MIT or Apache-2.0. Open-weight models are trained files you can download and run, but the licence may restrict use and training data and code are often not published, so many are not open source in the OSI sense.
Can I run AI on my own computer for privacy?
Yes. Ollama or LM Studio runs a model locally, and Open WebUI adds a browser interface. Prompts stay on your machine for local models, which is the main reason to run locally. The trade-offs are smaller models, hardware limits and that you maintain it. Quality depends on the model you can fit.
Is LM Studio open source?
No. The page lists LM Studio as proprietary: it is free to download but not open source, so you rely on the vendor. By contrast Ollama and llama.cpp are listed under the MIT licence. Licence statements for software were read on 2026-10-17, so check again before commercial use.
Is Open WebUI open source?
Not by the OSI definition. The page notes that Open WebUI has a custom licence with a branding clause, which limits rebranding. It is a self-hosted chat front end for Ollama and OpenAI-compatible APIs. Read the licence before commercial use, since field-of-use limits and branding clauses do not meet the open source definition.
Are open-weight models as good as ChatGPT or Claude?
For many everyday tasks they can be good enough, but the best closed models are often ahead on hard tasks, and the gap depends on the task. Larger models need much more memory, and licences vary by release. Do not choose from a single benchmark. Test on your own work.
D. Alternatives to social and messaging bots
Bots that read or post on social platforms should use the platform's official APIs. Observe rate limits and each platform's automation policy, disclose automated accounts where required, and keep a human approving public posts.
Workflow tools such as n8n can connect to official APIs. Treat them as automation with an AI step, and draft rather than post automatically.