Open weights vs closed AI models
What open weights really means, how it differs from open source, and the trade-offs between models you can download and models you can only use through an API.
Last updated Oct 1, 2026
At its core, every AI model is a very large file of numbers called weights. Those numbers are what the model learned during training. Whether a lab lets you download that file is one of the biggest dividing lines in AI, and every Modeldle card says which side a model is on.
What open weights means
An open weights model is one whose trained weights are published for anyone to download. With the file and enough hardware, you can run the model on your own computers, adapt it with further training (called fine-tuning), and keep using that exact version for as long as you like.
A closed model, also called proprietary or API-only, keeps its weights private. You can use it through the lab's own apps, its API, or cloud platforms it partners with, but the file itself never leaves the provider's servers.
Open weights is not the same as open source
The two phrases are often used as if they meant the same thing, but they do not. In software, open source means you get the source code and the freedom to use, study, change and share it. For AI, the Open Source Initiative's definition asks for more than weights: enough detail about the training data, plus the code used to train and run the model, for someone else to meaningfully rebuild it.
Most open weights releases stop short of that. Labs usually publish the weights and the code needed to run them, but not the training data, which may be enormous, licensed from others, or simply kept as a competitive secret. The license matters too:
- Some models use standard permissive licenses. gpt-oss-120b was released under Apache 2.0 and DeepSeek R1 under the MIT license, both of which allow commercial use with few conditions.
- Others use custom licenses with extra rules. Meta's Llama licenses, for example, include an acceptable use policy and require companies with very large user bases to request a separate license from Meta.
So "open" is a spectrum. Before you build a product on an open weights model, read its license.
Who releases what
This is not simply one group of labs against another. Several large labs publish both kinds of model:
| Lab | Open weights | Closed |
|---|---|---|
| OpenAI | gpt-oss-120b | GPT-6 Sol |
| Gemma 4 31B | Gemini 3.1 Pro | |
| Meta | Llama 4 Maverick | Muse Spark 1.3 |
| Alibaba (Qwen) | Qwen3.5 397B-A17B | Qwen3.8 Max |
| Mistral AI | Mistral Large 3 | Mistral Medium 3.1 |
Other labs lean fully one way. In the catalog, every model from DeepSeek, Moonshot AI (Kimi), Z.ai (GLM) and MiniMax has open weights, and so does NVIDIA's Nemotron 3 Ultra. Every Anthropic and xAI model in the catalog is closed.
How good are open models?
The gap has narrowed a lot. On the LMArena text leaderboard, the best closed model in the catalog, Claude Opus 5.5, scores 1512. The best open weights models, Kimi K3 at 1476 and GLM-5.3 at 1472, are within about 40 points. On the arena's scale, Opus 5.5's 36 point lead over Kimi K3 means people would pick it in roughly 55 percent of head-to-head votes: a real lead, but not a rout. The LMArena guide explains how those scores work.
The trade-offs
Privacy and control
Run an open model on your own hardware and your data never has to leave your machines, which matters for hospitals, banks, law firms and anyone with strict data rules. You also control the version. A closed model can be updated or retired on the provider's schedule, while a downloaded file stays the same forever. And you can fine-tune it on your own material to make it better at a narrow job.
Cost
Open models are often cheap even if you never touch a server, because many hosting companies compete to serve them. DeepSeek V4 Flash is listed at $0.15 per million input tokens and $0.60 per million output tokens. Running a model yourself, on the other hand, only saves money if you keep the hardware busy. Idle machines still cost money.
Hardware
This is where the idea of running everything at home meets reality. A model's memory needs grow with its parameter count. Kimi K3 has 2.8 trillion parameters. Even compressed to 4 bits per parameter, its weights alone take about 1.4 terabytes of memory, which means many data center GPUs working together. At the other end, Gemma 3 27B at 4 bits needs roughly 14 gigabytes, which fits on a single high-end consumer graphics card. Most people running models at home pick something in the tens of billions of parameters.
Closed models need none of this. You pay per request, and the provider handles servers, scaling and updates.
Safety and responsibility
A closed provider can monitor use, block misuse and fix problems for everyone at once. Once weights are published, they cannot be recalled, and safety training can be weakened by further fine-tuning. Supporters of openness argue that it lets far more researchers study, audit and improve models. Both points are part of an active debate, and labs weigh them differently.
In the quiz
Modeldle asks "Which of these can you download and run yourself?", hiding one open weights model among API-only ones. The lab name alone will not save you, since OpenAI, Google and Meta all appear on both sides. It pays to learn the model lines instead: gpt-oss, Gemma and Llama are open, while GPT, Gemini and Muse Spark are closed. Play today's quiz to put it to the test.