Open-weight vs closed vs open-source AI models: what is the difference?

Part of our AI models guide: 11 posts on choosing and running AI models.
You can use ChatGPT as much as you like, but you can never get a copy of the AI model behind it. You can download DeepSeek's model in an afternoon, but you cannot see exactly how it was made. With a rare few, such as Olmo, you get the model plus the full record of how it was built. Closed, open-weight and open source: three labels that get mixed up all the time, even by the companies that make AI.
This guide is current as of 19 November 2026. A model's weights are the billions of numbers it learned in training, saved as files: they are the model.
The restaurant comparison
Closed models, such as ChatGPT, Claude and Gemini, are like eating at the restaurant. You cannot take the dish home or see the recipe. Everything is cooked in their kitchen, meaning their servers. You order through their app or an API, a way for your software to send the model questions over the internet and get answers back. APIs charge per token, the smallest unit of text an AI model reads and writes, often a short word or part of a longer one (see cost per token, explained).
Open-weight models, such as DeepSeek, Qwen, Llama and OpenAI's gpt-oss, hand you the finished dish to take home. You download the weights and run the model in your own kitchen: your own servers, fitted with GPUs, the specialised chips that run AI models. You can add your own spices through fine-tuning: training the model a little further on your own examples so it suits your work. You do not get the recipe (the code that trained it) or the list of ingredients (the data it learned from).
Open source, in the strict sense, is the dish plus the full recipe (the training code) plus where every ingredient came from (the training data). A skilled cook could make it again from scratch. Very few models qualify; Olmo, from the US non-profit lab Ai2, is one.
The licence is the house rules printed on the box: the legal terms that say what you may do with the model.
Closed vs open-weight vs open source
At a glance | Closed | Open-weight | Open source (strict sense) |
|---|---|---|---|
What you get | Answers, through an app or API | The weights, to run where you like | Weights, complete code and training-data details |
Examples | ChatGPT, Claude, Gemini | DeepSeek, Qwen, Kimi, Llama, gpt-oss | Olmo |
How you pay | Per token, or a monthly plan | Free to download; you pay for the GPUs, or per token to a company that runs it for you | Same as open-weight |
Control and privacy | What you type goes to the maker's servers | On your own servers, what you type stays with you | Same as open-weight, and you can check how it was built |
Who it suits | Teams starting out; tasks that need top quality | Steady heavy use, private data, fine-tuning | Researchers and auditors |
What open source means for AI
The Open Source Initiative (OSI) is the non-profit that has defined open-source software since 1998. It published its Open Source AI Definition 1.0 on 28 October 2024, still the current version. The definition asks for the freedom to use a model for any purpose without asking permission, study how it works, change it and share it. That takes three things together:
- The weights, under terms that grant those freedoms.
- The complete code used to train and run the model.
- Enough detail about the training data that a skilled person could build a substantially equivalent model.
Weights alone never qualify. When the OSI tested its definition in 2024, an earlier Olmo passed and Meta's Llama 2 did not, though the OSI stresses these were not certifications. K2 Horizon, from Abu Dhabi's MBZUAI university (see Arabic models from the Gulf), publishes its weights, training data and code. Switzerland's Apertus 1.5 publishes its weights and code with full details of its training data. Both describe themselves as fully open, though the OSI has not reviewed them.
Openness is per model, not per company
Most big labs do both: a closed flagship (see the big US models), plus open weights for other models.
Maker | Open weights | Closed |
|---|---|---|
OpenAI | gpt-oss-120b and gpt-oss-20b | GPT-6.2 Astra and Sol |
Gemma 4 | Gemini 4 | |
Alibaba | Qwen3.8 models | Qwen4-Max |
Meta | Llama 4 and Muse Spark Open | Muse Wonder |
Open weights are not only a US and Chinese story: India's Sarvam, Japan's LLM-jp and Korea's LG publish open-weight models too (see Indian models and Japanese and Korean models).
The licence: house rules on the box
Permissive licences, such as MIT and Apache 2.0, let you use, change and resell the model with few conditions. Custom licences add some: a name to display, a revenue level above which you need a separate deal, or a request form before you can download.
Model | Licence | Download | Sell access to it as a service? | Full recipe? |
|---|---|---|---|---|
DeepSeek V4.1 Flash | MIT | Open to all | Yes | No |
Kimi K3.1 Pro | Custom | Open to all | Yes, until your company group's revenue tops $20M in any 12 months; then a separate deal | No |
Muse Spark Open | Muse Model License | After Meta approves your request | Yes, but above $20M group revenue in any 12 months, tell Meta and sign a deal | No |
Mistral Large 4 | Modified MIT | Open to all | Only with a separate commercial licence from Mistral, at any revenue | No |
Olmo 3.1 | Apache 2.0 | Open to all | Yes | Yes |
In-house use is exempt under the Kimi, Muse and Mistral licences, so their extra conditions apply mainly if you sell access to the model as a service. Terms change between versions, so read the licence every time. For DeepSeek, Qwen and Kimi in depth, see Chinese models; for Mistral and Apertus, European models.
Which should a business choose?
- Quality: closed models still lead. On the Intelligence Index from Artificial Analysis, an independent firm that tests AI models, Claude Fable 5.2 scores 65 and the top eleven are all closed. The best open-weight model, Kimi K3.1 Pro from China's Moonshot AI, scores 46 in 12th place, 19 points behind.
- Price: open-weight tokens cost far less. On OpenRouter, a single API for hundreds of models, closed models averaged $0.52 per million tokens in October 2026 and open-weight models $0.084, about a sixth of the price, by OpenRouter's own data. That is why cheap open-weight models carry so much traffic (see the most popular AI models).
- Control: with open weights on your own servers, what you type stays there. OpenAI and Anthropic generally keep API data for up to 30 days unless you sign a special agreement for them to keep none.
- Effort: someone has to run it. gpt-oss-20b fits in 16 GB of memory, so a well-equipped laptop can run it. Kimi K3.1 Pro's weights are about 1.56 TB, so the smallest single server that holds them has eight NVIDIA B300 GPUs. See our GPU sizing guide.
There is a middle road: inference providers run open-weight models on their own servers and sell the answers per token (inference is the industry word for running a model). That is like eating the take-home dish at a second restaurant: you need no kitchen, but what you type still goes to their servers (see how inference providers sell tokens).
A simple rule: prove the idea on a closed API, then move to open weights when volume is steady, data must stay in-house or you need a fine-tuned model. Read the licence first.
Your own kitchen for open weights
Taking a large open-weight model home means you need a kitchen. FlexiCloud arranges whole 8-GPU servers (H100, H200, B200 and B300, with GB300 NVL72 racks on request) from verified GPU server operators in North America, the UK, Europe, India, Asia Pacific and the Gulf. Terms run a year or more, billed upfront. Tell us the model, how much text it must read at once and how busy it will be, and we will size the server, or see the GPU servers we source.
Where these figures come from
- Definitions: OSI, Open Source AI Definition 1.0 (28 October 2024) and its FAQ; OSI opinion on the Llama licence (18 February 2025).
- Licences and sizes, checked on Hugging Face on 19 November 2026: DeepSeek V4.1 Flash, gpt-oss-120b, Qwen3.8-27B, Kimi K3.1 Pro, Muse Spark Open, Mistral Large 4, Llama 4, Sarvam 105B, LLM-jp 4.1 and K-EXAONE 2.0.
- Fully open models: Ai2 Olmo, K2 Horizon (3 September 2026) and CSCS on Apertus 1.5 (24 July 2026).
- Open and closed by maker: TNW on Qwen4-Max (16 November 2026) and The Register on Gemma 4 (2 April 2026).
- Quality: Artificial Analysis leaderboard, Intelligence Index v5 (19 November 2026).
- Prices: OpenRouter, October 2026 data (OpenRouter's own data).
- Data retention: OpenAI, your data and Anthropic privacy centre (updated 1 July 2026).