Open-Weight vs Closed AI Models: What the Difference Means for You

An open-weight AI model is one whose trained parameters, known as weights, are published. Anyone can then download the model and run it on their own hardware. A closed model stays on servers run by its developer or the developer's cloud partners. Users reach it through apps or a paid programming interface, called an API. The difference affects where data is processed, what the model costs to use and how much control the user has. It does not, by itself, make a model open source.

What is an open-weight AI model?

Weights are the numbers a neural network learns during training. The Open Source Initiative (OSI) describes open weights as the finished weights and biases a network ends up with after training. Once shared, OSI says, they let other people fine-tune, adapt or deploy the model themselves.

As of October 6, 2026, open-weight releases include:

• DeepSeek-V4.1-Flash
• Qwen3.8-27B
• Mistral Large 3
• Google's Gemma 4 family
• Meta's Llama 4 and Muse Glimmer
• OpenAI's gpt-oss models

Mistral's documentation also lists a newer Mistral Large 4, version 26.10, which it describes as open-weight. The line does not run neatly between companies. OpenAI and Google sell access to closed flagship models and also publish smaller open-weight ones. Meta publishes weights for Llama 4 and Muse Glimmer. But a September 2026 Meta developer post lists its own API and partner clouds as the ways to use Muse Spark 1.3.

Hugging Face is the site where DeepSeek, Qwen, Meta and OpenAI post their weights. It listed more than 3.1 million models that day.

Is open-weight the same as open source?

No. OSI's Open Source AI Definition says users must be free to use, study, modify and share a system. It requires three things:

• The model's weights.
• The complete code used to train and run it.
• Enough detail about the training data that a skilled person could build a broadly equivalent system.

OSI says open weights on their own leave out the training code and the data details.

Licenses also differ:

• Standard open licenses. DeepSeek-V4.1-Flash is released under the MIT License. Qwen3.8-27B, Meta's Muse Glimmer 30B and OpenAI's gpt-oss-20b use Apache 2.0. So do Mistral Large 3 and Mistral Small 4, according to Mistral's documentation.
• Custom licenses. Llama 4 comes under Meta's own Llama 4 Community License. Under it, companies whose products had more than 700 million monthly active users when Llama 4 was released must request a separate license from Meta. It also requires products that include the model to display a Built with Llama notice.

Is DeepSeek open source? Its weights are openly licensed under MIT. Whether the whole system meets OSI's stricter definition depends on the training code and data information released with the weights, not on the license alone. OSI argued in February 2025 that Meta's earlier Llama 3 licenses were not open source because of their restrictions on use.

Can I run AI locally on my PC?

Yes, provided the model fits in the machine's memory. OpenAI's model card for gpt-oss-20b says that model runs within 16GB of memory. It says the larger gpt-oss-120b fits on a single 80GB graphics processor (GPU) such as Nvidia's H100.

Models can be shrunk through quantization, which stores each weight with fewer bits. Google's Gemma 4 documentation gives figures for the largest Gemma 4 model, which has 31 billion parameters. It says loading that model takes about 69.9GB of memory at 16-bit precision but about 17.5GB at 4-bit. The smallest takes about 2.9GB at 4-bit. Google notes that these estimates cover the weights only and that long prompts need more. Hugging Face pages for gpt-oss-20b and Qwen3.8-27B point to tools such as Ollama, LM Studio and llama.cpp for running models on a personal computer.

Not every open-weight model is PC-sized. DeepSeek's model card gives DeepSeek-V4.1-Flash 552 billion backbone parameters, roughly 26 times the 21 billion in gpt-oss-20b.

Are open models as good as closed ones?

In the latest Stanford figures, closed models were ahead. The 2026 AI Index reports that, as of March 2026, the top closed model led the top open model by 3.3%. That lead was up from 0.5% in August 2024, the report says. It also reports that six of the top ten models on the Arena leaderboard were closed.

Price is less clear-cut. DeepSeek's hosted service lists DeepSeek-V4.1-Flash at $0.30 per million uncached input tokens and $1.20 per million output tokens at peak hours. Tokens are small units of text. The service lists half that off-peak. OpenAI lists its closed flagship, GPT-6 Astra, at $10 and $50. It lists its lowest-priced GPT-6 model, Luna, at $0.10 and $0.50. The three are not equivalent in ability, so this is not a like-for-like comparison. Running a model locally removes per-token fees altogether, though the user pays for hardware and electricity instead.

The bottom line

Open-weight models can be customized and run on private hardware, and the smaller ones fit on an ordinary PC. Closed models require no hardware of the user's own. In Stanford's March 2026 data, they held a lead at the top end. The label matters less than the license attached to each model. The license sets out what users may do with it.