Closed AI models, such as Claude, GPT and Gemini, run only on their makers’ servers. You rent them through an app or an API. Open-weight models, such as DeepSeek, Qwen, Gemma and gpt-oss, can be downloaded and run on hardware you control. Closed models still lead on the hardest tasks and are easier to start with. Open ones win on control, privacy and cost at scale.
There is a catch in the words. Most models called open source are really open-weight: you get the finished model, not the data and code that made it. And the license, not the download button, decides what you may do with it.
- Closed models run on the vendor’s servers. Open-weight models run wherever you put them.
- Open weights are not open source. Most releases share the trained model, not the training data and code.
- Closed models lead on peak quality and convenience. Open ones win on control, privacy and cost at high volume.
- The gap is small: in July 2026, one close observer put it at three to five months.
- Read the license. Apache 2.0 and MIT are simple, while custom licenses can add user caps, revenue triggers and naming rules.
Open weights vs open source: what is the difference?
Think of a cake. Open weights hand you the finished cake: you can serve it, slice it and add your own frosting. Open source, in the full sense, also hands you the recipe and the shopping list, so you could bake it again and change how it is made.
The Open Source Initiative, which maintains the standard definition of open-source software, has published an Open Source AI Definition. Version 1.0 requires the freedom to use, study, modify and share an AI system for any purpose, without asking permission. It also requires three components:
Data information: enough detail about the training data that a skilled person could build a substantially equivalent system
Code: the complete source code used to train and run the system
Parameters: the weights themselves
By that bar, a model released as weights alone, under a license that limits who may use it, is open-weight but not open source. Fully open releases do exist. The Allen Institute for AI publishes its Olmo 3 models with their training data, training code, intermediate checkpoints and logs.
| Term | What you get | Examples, as of September 2026 |
|---|---|---|
| Closed | Access through an app or API only | Claude, GPT-6, Gemini, Meta’s Muse Spark |
| Open-weight, custom license | The model to download, with the maker’s conditions | Llama 4, Kimi K3, GLM-5.3 |
| Open-weight, permissive license | The model under Apache 2.0 or MIT | Gemma 4, Qwen3.8 27B, gpt-oss, DeepSeek V4-Pro |
| Fully open | Weights plus training data, code and checkpoints | Olmo 3 from AI2 |
The trade-offs at a glance
| Closed models | Open-weight models | |
|---|---|---|
| Peak quality | Hold the top spots on the hardest tasks | Months behind at the top; small sizes trail further |
| Getting started | Sign up and go | Choose, host and update the model yourself, or pay a host |
| Privacy | Your data goes to the vendor, under its terms | Can stay entirely on hardware you control |
| Cost | Pay per token or per seat | Hardware or hosting; can be cheaper at high, steady volume |
| Control | The vendor can change or retire a model | Keep a version forever and fine-tune it |
| Safeguards | Built in and run by the vendor | Whatever ships with the model; you add the rest |
| Terms | The vendor’s terms of service | The model’s license, from Apache 2.0 to custom |
Each row hides a real decision:
Control. With open weights, the model you tested is the model you run, for as long as you like. Closed models retire: Anthropic, for example, publishes a date for each model before which it will not be retired, as early as October 2026 for Claude Haiku 4.5.
Customization. You can fine-tune open weights on your own data. Closed vendors offer this in limited forms, if at all: Anthropic’s API, for one, does not currently offer fine-tuning. Our guide to fine-tuning vs RAG vs prompting explains when that customization actually pays off.
Privacy. An open model can run on a laptop or inside your own cloud account, so sensitive data never reaches a third party. Our guide to running AI models locally shows how.
Cost. The weights are free, but hardware, power and engineering time are not. For light use, paying per token is usually cheaper. At high, steady volume, running your own model can win.
Size. The strongest open models are enormous. Kimi K3 has 2.8 trillion parameters and DeepSeek V4-Pro has 1.6 trillion, which means data center hardware, not a laptop. Smaller open weights, covered in our guide to small language models, fit on a laptop instead.
License gotchas: free to download is not free to use
An open-weight model comes with a license, and the license decides what you may do. Two standard licenses cover many popular models:
Apache 2.0: Gemma 4, Qwen3.8 27B and gpt-oss.
MIT: DeepSeek V4-Pro.
Both are permissive. They allow commercial use and changes, as long as you keep the license and copyright notices. Custom licenses are where the surprises live:
Llama 4 Community License (Meta). Products with more than 700 million monthly active users must request a license, which Meta may grant at its sole discretion. You must display “Built with Llama”, start the name of any model you build from it with “Llama”, and follow Meta’s acceptable use policy.
Kimi K3 License (Moonshot AI). It is based on MIT, with two triggers. A company whose revenue, with its affiliates, tops $20 million over any 12 months needs a separate agreement before commercial use. Products with more than 100 million monthly users, or $20 million in monthly revenue, must display “Kimi K3” prominently. Internal use is exempt.
GLM-5.3 (Z.ai). It ships under its own license rather than a standard one, so read it in full before you build on it.
The 2026 landscape: who is on each side?
The closed side is led by the big American labs. Anthropic sells Claude Opus 5.5 and Fable 5.1, OpenAI sells GPT-6 Astra, Sol and Luna, and Google sells Gemini 3.8 Flash and 3.1 Pro. Meta, known for its open Llama models, launched its newer Muse Spark models through an API and its own apps.
The strongest open-weight models now come mostly from Chinese labs: DeepSeek, Moonshot AI, Alibaba’s Qwen team and Z.ai. The American open releases are smaller, such as Google’s Gemma 4 and OpenAI’s gpt-oss.
- April 2026DeepSeek releases V4, with V4-Pro at 1.6 trillion parameters under the MIT license
- July 2026Meta launches Muse Spark 1.1 through its API in public preview, with no weights announced
- July 2026Moonshot AI releases the weights of Kimi K3, a 2.8-trillion-parameter model, under a custom license
- August 2026Alibaba’s Qwen3.8 arrives, including the Apache 2.0 Qwen3.8-27B
- September 2026OpenAI adds GPT-6 Sol and Luna to its closed GPT-6 family, alongside Astra
How far behind is open? Nathan Lambert, who tracks open models closely, estimated in July 2026 that the gap had narrowed from a debated six to nine months to about three to five. At the time, Kimi K3 ranked third on Artificial Analysis’s Intelligence Index, behind only Claude Fable and GPT-5.6 Sol.
Leaderboards shift with every release, and each lab picks the tests it reports. Our guide to AI benchmarks explains how to read the numbers both camps quote.
Which should you choose?
Choose a closed model when:
You need the best possible answer on hard, multi-step work
You want to start today with no infrastructure
Your volume is low or unpredictable
Choose an open-weight model when:
Data must stay on hardware you control
You need the same model version for years
You run high, steady volume
You want to fine-tune on your own data
You do not have to pick a side forever. Prototype with a closed model, then move stable, high-volume jobs to an open one once you know what good looks like. Our guide on how to choose an AI model shows how to test candidates on your own tasks.
FAQ
Is Llama open source?
Not by the Open Source Initiative’s definition. Llama’s weights are downloadable, but its license adds conditions, such as asking Meta for permission above 700 million monthly users and following an acceptable use policy. Open-weight is the accurate term.
Can I use open-weight models commercially?
Usually, but the license decides. Apache 2.0 and MIT models, such as Gemma 4, gpt-oss and DeepSeek V4-Pro, allow it. Custom licenses may add attribution rules, user caps or revenue triggers.
Are open models less safe?
They move the responsibility. A closed vendor applies its own safeguards on its servers. With open weights, whoever runs the model decides which safeguards to add, and anyone can change the model.
Are open-weight models cheaper?
The weights are free, but running them is not. For light use, a closed model’s pay-per-token price is usually cheaper than buying hardware. At high, steady volume, running your own model can cost less.
What does open weights mean?
It means the trained model is published for anyone to download, run and modify. It does not mean the training data or the training code is public.
- Closed models are rented through an API or app. Open-weight models run wherever you choose.
- Open weights are not open source: fully open releases like Olmo 3 also share data and code.
- Closed models lead at the very top, but the best open models trail by months, not years.
- The license decides what you may do. Check user caps, revenue triggers and naming rules.
- Many needs are best met with both: closed for hard problems, open for private or high-volume work.
Next, learn how to run AI models on your own machine, or see how to choose an AI model for the job.
- The Open Source AI Definition 1.0, Open Source Initiative
- Olmo, Allen Institute for AI
- Llama 4 Community License Agreement, Meta, April 2025
- Kimi K3 model card and license, Moonshot AI, July 2026
- GLM-5.3 model card, Z.ai
- DeepSeek-V4-Pro model card, DeepSeek
- Gemma 4 31B model card, Google
- Qwen3.8-27B model card, Qwen
- gpt-oss in the Ollama library, Ollama
- Introducing Muse Spark 1.1, Meta, July 2026
- Kimi K3: the open weights escalation, Nathan Lambert, Interconnects, July 2026
- Models overview and Glossary, Anthropic, September 2026
- OpenAI’s GPT-6 Sol and GPT-6 Luna now available, GitHub, September 2026
