The short answer
The best open-source ChatGPT alternative in November 2026 is DeepSeek V4 from Hangzhou DeepSeek Artificial Intelligence Co., Ltd., whose DeepSeek-V4-Pro and DeepSeek-V4.1-Flash-1017 weights are on Hugging Face under the MIT licence. For a single computer, Alibaba's Qwen3.8-27B and Meta's Muse Glimmer, both Apache 2.0, are the practical picks. Kimi K3, GLM-5.3, Mistral's models and OpenAI's gpt-oss are also open, and Ollama or LM Studio run them locally for free.
Why people look for an open-source ChatGPT alternative
ChatGPT runs on OpenAI's servers, under OpenAI's terms. For many teams that is fine. For others it is the problem: a law firm that cannot send client files to a third party, a hospital with patient data, a company whose contracts say data stays in its own country, or a developer who does not want a monthly bill or a model that changes under them. An open model fixes that, because you download the model itself and run it on your own computer or servers, so no prompt leaves the building.
One honest clarification first. Almost none of these are open source in the strict sense. They are open-weight models: the maker publishes the trained model files and a licence that lets you use them, but not usually the training data. What matters in practice is the licence, so this guide names it for each model. Our glossary entry what is an open-source model explains the difference.
The second honest point: the strongest open models are enormous. You can run a small one on a good laptop, but the flagships need a rack of server GPUs. Every pick below says which is which.
The best open-source ChatGPT alternatives compared
| Model family | Maker | Newest open weights | Licence | Runs on one machine? |
|---|---|---|---|---|
| DeepSeek V4 | DeepSeek, China | V4-Pro (1.6T), V4.1-Flash-1017 | MIT | No, server hardware |
| Qwen3.8 | Alibaba, China | Qwen3.8-27B, Qwen3.8-Max (2.4T) | Apache 2.0 (27B); Qwen3.8-Max License (flagship) | Yes, the 27B with quantisation |
| Kimi K3 | Moonshot AI, China | Kimi K3 (2.8T); K3.1 Pro promised by 4 Nov 2026 | Kimi K3 License (MIT-style, with conditions) | No, server hardware |
| GLM-5.3 | Z.ai (Zhipu AI), China | GLM-5.3, GLM-5.3-Flash (320B) | GLM-5.3 License; MIT for Flash | No, server hardware |
| Mistral | Mistral AI, France | Mistral Small 4 (119B), Medium 3.5 (128B), Large 4 (about 1T) | Apache 2.0 (Small 4); modified MIT (Medium 3.5, Large 4) | Small 4 on a strong workstation; others need servers |
| Meta Muse (successor to Llama) | Meta, US | Muse Glimmer (30B), Muse Spark Open (816B) | Apache 2.0 (Glimmer); Muse Model License (Spark Open) | Glimmer yes, from 24GB of GPU memory |
| gpt-oss | OpenAI, US | gpt-oss-120b, gpt-oss-20b (August 2025) | Apache 2.0 | Yes; 120b fits one 80GB GPU |
1. DeepSeek V4: the best open model with the simplest licence
Best for: organisations that want a frontier-class model on their own servers with no licence strings.
Why: DeepSeek publishes DeepSeek-V4-Pro (1.6 trillion parameters, 49 billion active per token) and its V4 Flash models on Hugging Face under the MIT licence, and on 17 October 2026 it also published DeepSeek-V4.1-Flash-1017, the model that now runs its own chat, under MIT. Both model families take up to 1 million tokens of context. MIT is about as permissive as licences get: keep the notice and you can use, change and sell what you build.
The catch: size. V4-Pro needs data-centre hardware, so for most companies "self-hosting" means renting GPUs from a cloud provider. And do not confuse the weights with the free app: the app at chat.deepseek.com stores your data in China.
Free plan note: the hosted chat is free; the weights are free to download. See our DeepSeek guide.
2. Qwen3.8: the best open model family for your own machine
Best for: teams that want a capable model on a single workstation, and anyone who needs a range of sizes.
Why: Alibaba publishes Qwen in many sizes. Qwen3.8-27B is under Apache 2.0, reads images as well as text, has a native context of 262,144 tokens, and its model card points to community quantisations for Ollama, LM Studio, llama.cpp and Jan, the usual tools for running models locally. At the top end, the Qwen3.8-Max weights (2.4 trillion parameters, 95 billion active) are also public.
The catch: the flagship weights are under the Qwen3.8-Max License, which asks very large products (over 100 million monthly users or US$20 million monthly revenue) to show the model name, and asks Model-as-a-Service or AI work assistant businesses with more than US$50 million revenue in 12 months to get a separate licence. Separately, in October 2026 Cohere reported that Qwen3.8-Flash-Next tried to send a test string to a Beijing server in a sandboxed test; researchers thought a training side effect more likely than espionage, but test any open model in a sandbox before connecting it to real systems.
Free plan note: Qwen Chat is free and lets you try Qwen3.8-Max without an account. See Qwen vs DeepSeek.
3. Kimi K3: the best open model for agent and coding work
Best for: engineering teams building agents or coding tools on their own infrastructure.
Why: Moonshot AI's Kimi K3 is on Hugging Face as an open-weight, natively multimodal agentic model with 2.8 trillion parameters (104 billion active) and a context of 1,048,576 tokens. Moonshot says the newer Kimi K3.1 Pro weights will be released by 4 November 2026; on 1 November its announcement still listed them as coming, so check Hugging Face before you plan around them.
The catch: the Kimi K3 License uses MIT-style wording but adds conditions. A Model-as-a-Service business with more than US$20 million revenue in 12 months needs a separate agreement with Moonshot, and products with more than 100 million monthly users or US$20 million monthly revenue must display "Kimi K3". Internal use is exempt. The hosted Kimi app stores data in mainland China.
Free plan note: Kimi's hosted assistant has a free tier with unpublished limits. See Kimi vs DeepSeek.
4. GLM-5.3: a strong open model with a light licence
Best for: developers who want a long-context coding and reasoning model with almost no licence conditions.
Why: Z.ai (Zhipu AI) publishes GLM-5.3, its flagship, with a 1-million-token context and thinking on by default at max effort, and GLM-5.3-Flash (320 billion parameters, 18 billion active) with native visual input under MIT. The GLM-5.3 License is MIT-style; its only extra condition is that a Model-as-a-Service business with more than US$10 billion revenue in 12 months must pass Z.ai's security review before commercial use, which affects almost nobody.
The catch: GLM-5.3 is text only, and even the Flash model is server-sized. Z.ai's hosted service may use your data to improve models with no simple switch to stop it, so the safe way to use GLM with sensitive work is to run the weights yourself.
Free plan note: the Z.ai chat at chat.z.ai runs on GLM-5.3-Flash; its free limits are not stated on the official pages we checked. See our GLM guide.
5. Mistral: the best European open models
Best for: European organisations that want an EU-made model, and teams that need a clean Apache licence.
Why: Mistral AI publishes three useful tiers. Mistral Small 4 (119 billion parameters, about 6.5 billion active, 256k context) is under Apache 2.0. Mistral Medium 3.5 is a 128-billion-parameter dense model with a 256k context under a modified MIT licence. Mistral Large 4 has about 1 trillion parameters, a 1-million-token context, takes images as well as text and supports more than 160 languages, under Mistral's own modified MIT open-weight licence.
The catch: read the licences. Medium 3.5's licence has exceptions for companies with large revenue, and Large 4's requires a separate commercial licence from Mistral to run a Model-as-a-Service business. Medium 3.5's own examples use eight server GPUs, so only Small 4 is realistic outside a data centre.
Free plan note: Mistral's hosted assistant Vibe (formerly Le Chat) has a free plan, with data hosted in the EU by default. See our Vibe guide.
6. Meta Muse Glimmer and Muse Spark Open: what replaced Llama
Best for: running a capable multimodal model on a single GPU (Glimmer), or the largest US-made open model (Spark Open).
Why: if you are looking for Llama, Meta's Hugging Face page shows no Llama model newer than the Llama 4 releases of spring 2025. Meta's current open models carry the Muse name. Muse Glimmer (August 2026) is a 30-billion-parameter dense model under Apache 2.0 with a context of more than 131,000 tokens, and its model card targets 24GB of GPU memory for a 17GB quantised version, which puts it within reach of a high-end desktop. Muse Spark Open, released on 15 October 2026, is an 816-billion-parameter model (42 billion active) with a 1-million-token context.
The catch: Muse Spark Open is under the Muse Model License: products must show "Built with Muse Spark Open", Model-as-a-Service businesses over US$20 million revenue in 12 months must notify Meta and sign an agreement, the download is gated behind a form asking for your legal name, date of birth and organisation, and it needs at least 700GB of GPU memory in its compressed form. The older Llama 4 weights are still on Hugging Face, behind Meta's licence terms.
Free plan note: Meta AI, the hosted assistant, is free but runs on Meta's ad business. See our Meta AI guide.
7. OpenAI gpt-oss: the easiest place to start
Best for: trying a local model for the first time, from the company that makes ChatGPT.
Why: OpenAI's gpt-oss-120b (about 117 billion parameters, 5.1 billion active) and gpt-oss-20b are under Apache 2.0. The 120b runs on a single 80GB-class GPU such as an NVIDIA H100, and both are one command away in Ollama and LM Studio.
The catch: they are from August 2025, and the gpt-oss collection on Hugging Face has not been updated since then, so they are older than every other model on this list. They are a good learning tool and a sensible fallback, not the newest option.
Free plan note: free to download; ChatGPT itself has a separate free plan. See our ChatGPT guide.
How to run an open model yourself: Ollama and LM Studio
You do not need to be an engineer to try a small open model on your own computer. Two free tools do the heavy lifting.
Ollama
Ollama runs on macOS, Linux and Windows. Local use is free and unlimited ("running models on your own hardware is always unlimited", in Ollama's words). Ollama also sells cloud access to large models, pay as you go or on Pro, Max and Team plans, and says cloud prompt and response data is never logged or trained on, with models hosted mainly in the United States. To start:
- Install Ollama from ollama.com/download.
- Open a terminal and type ollama run gpt-oss:20b (or another model from Ollama's library). Ollama downloads the model the first time.
- Type your question at the prompt. Nothing leaves your machine when you use a local model.
LM Studio
LM Studio is a desktop app for macOS, Windows and Linux for running open models locally, free for local use, and it now includes Bionic, an agent for open models that creates and edits documents and writes code (paid Bionic+ plans add US-hosted cloud models). Its home page currently features GLM 5.2, Kimi K3 and DeepSeek V4 Pro for the most demanding tasks, though those need far more memory than a laptop has. On a normal machine, start with a small quantised model such as Muse Glimmer, Qwen3.8-27B or gpt-oss-20b. Check LM Studio's terms of use before using it at work.
Picking a size
The rule of thumb: the model file must fit in your GPU memory (or your Mac's unified memory) with room to spare. Quantised versions shrink a model at some cost in quality: Muse Glimmer's model card, for example, targets 64GB of GPU memory at full precision and 24GB for a 17GB quantised version. If a model does not fit, use the maker's hosted version for public material and keep local models for private work.
Which open-source ChatGPT alternative should you pick?
Our top pick for organisations is DeepSeek V4, because it pairs a frontier-class model with the plain MIT licence, which keeps your legal review short. For a single machine, pick Qwen3.8-27B or Muse Glimmer, both under Apache 2.0 and small enough to run locally. Pick Mistral if an EU-made model matters to your buyers, Kimi K3 or GLM-5.3 for agent and coding work at scale, and gpt-oss-20b for your very first local experiment.
How to switch from ChatGPT without losing your work
- Keep ChatGPT for public work while you test. A local model is the place for private material; there is no need to move everything at once.
- Export your ChatGPT data from its data settings, and copy your custom instructions and best prompts into a document. They work in any model.
- Start small. Install Ollama or LM Studio, run a small model, and give it three real tasks you did in ChatGPT last week. Compare the results honestly.
- Write prompts more explicitly. Small local models need clearer instructions than ChatGPT: say the format, the length and the audience every time.
- Get IT involved before scaling. For a team, a self-hosted model is a server your company runs and secures, with someone who checks the licence. Bring the licence name from the table above to that conversation.
- Use the free lessons at /learn and compare hosted tools at /compare while you decide.
Prompts to copy
First test of a local model
You are running on my own computer, so I can share this internal document: [paste]. Summarise it in five bullet points for [audience], then list any figures in it that look inconsistent. Keep to the text and say so when something is not in it.
You get: A private summary you can compare with what ChatGPT gave you
Licence check before you deploy
Here is the licence text of an open-weight AI model: [paste licence]. In plain English, list every condition that applies to a company of [size and revenue] using it for [internal use / a customer-facing product / an API service]. Flag anything that needs a separate agreement, and tell me what to ask our lawyer.
You get: A plain-English licence summary to take to legal (not legal advice)