# What are open-weight models? Open weights vs open source

> An open-weight model publishes its trained parameters, the weights, so anyone can download and run it. Open source asks for more: the training code and details of the training data. Most models called open, including gpt-oss, Qwen, DeepSeek, Gemma, Llama and Mistral, release the weights under a license, and that license sets what you may do with them.

Mindket editors · Published 2026-11-06 · Updated 2026-11-06 · https://mindket.com/learn/open-weight-models/what-are-open-weight-models

## In short

- Find a model's weights, license and card on its Hugging Face page before you download anything.
- Call a release open source only when it also ships training code and names its training data.
- Read the license field as the rule for your use, whatever the headline says.
- Compare sizes across families on their cards, and note whether a number counts all the parameters or only the active ones.

## Before you start

You need a browser. The model pages in this lesson open without an account, except Meta's Llama 4 page. Hugging Face [gates](https://huggingface.co/docs/hub/models-gated) that one: you log in and fill in a form that includes accepting Meta's license, and the Hub lists the access as manual approval.

A model's weights are the numbers it learned during training, billions of them, stored in files. They decide how the model answers. Every other file in a download, such as the tokenizer, the settings and the license, only helps your software load and use them.

## Inspect a release before you download it

Open one release and look at what it holds. The example is OpenAI's gpt-oss-20b, and the steps work the same on any model page. The file names in the steps come from the list that Hugging Face's API returned on 6 November 2026.

The card also gives two commands. The first pulls the model through Ollama, which [the Ollama tutorial](https://mindket.com/learn/local-ai/ollama-tutorial) sets up. The second downloads the weights with the Hugging Face command line tool.

````bash
# Ollama, from the model card
ollama pull gpt-oss:20b
ollama run gpt-oss:20b

# Download the weights with the Hugging Face CLI, from the model card
huggingface-cli download openai/gpt-oss-20b --include "original/*" --local-dir gpt-oss-20b/
````

1. Open [huggingface.co/openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b). The page is the model card, a README.md file with a block of settings at its top.
2. Find the license in that block. The card says `license: apache-2.0`.
3. Select the Files and versions tab. The files that end in `.safetensors` are the weights.
4. Open `config.json`. It holds the model's settings. The tokenizer files and `chat_template.jinja` next to it tell your software how to turn text into numbers and back.
5. Open `LICENSE` and `USAGE_POLICY`. The first is the Apache License 2.0. The second is two sentences and asks you to follow applicable law.
6. Look for what is missing. The repository holds no training data and no training code.
7. Run the model if you have the memory. The card says gpt-oss-20b runs within 16GB.

## Weights, code and data: what a release can hold

What you just inspected is weights plus the files that load them. Two other things go into building a model, and neither sat in that folder: the training code, the programs that ran the learning, and the training data, the text it learned from. The weights let you run a model and fine-tune it. Rebuilding it from scratch needs the other two.

That gap is the argument over open weights and open source. [Wikipedia's article on open-source AI](https://en.wikipedia.org/wiki/Open-source_artificial_intelligence) says the Open Source Initiative's definition asks for the full code to process the data, train the model and run it. It also asks for details about the training data, which need not be the data itself. A model that releases only the weights is open-weight, and the article notes critics call a weights-only release sold as open source "openwashing". It adds that Meta's description of Llama as open source has been disputed because of Llama's license.

Ai2's Olmo sits at the other end. The [Olmo 3.1 32B Think card](https://huggingface.co/allenai/Olmo-3.1-32B-Think) names the datasets it was trained on, Dolma 3 and Dolci, and says Ai2 is releasing all code, checkpoints and training details, with logs coming soon.

For choosing a model to run, the difference matters less than it sounds, because the download is the same. It matters when you must audit what a model learned or reproduce it, and when a company policy allows only open source.

**What three releases hold, by their own cards**

| | gpt-oss-20b | Llama 4 Scout | Olmo 3.1 32B Think |
| --- | --- | --- | --- |
| Weights to download | ✓ Yes | ◐ Partly | ✓ Yes |
| Training data named on the card | ✗ No | ◐ Partly | ✓ Yes |
| License | Apache 2.0, plus a usage policy | Llama 4 Community License | Apache 2.0 |

Sources: [Hugging Face: openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) (read 2026-11-06), [Meta: Llama 4 model card](https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md) (read 2026-11-06), [Hugging Face: meta-llama/Llama-4-Scout-17B-16E-Instruct](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) (read 2026-11-06), [Hugging Face: allenai/Olmo-3.1-32B-Think](https://huggingface.co/allenai/Olmo-3.1-32B-Think) (read 2026-11-06)

> Partly means with a condition. Llama 4 Scout's files on Hugging Face are gated behind a login, a form and manual approval. Its card describes the training data in words, as "a mix of publicly available, licensed data and information from Meta's products and services".

## The families you will meet

This course follows six families: gpt-oss from OpenAI, Qwen, DeepSeek, Llama from Meta, Gemma from Google DeepMind and Mistral. Some publish a ladder of sizes. Gemma 4's card lists five, E2B, E4B, 12B, 26B A4B and 31B. Others publish two: the gpt-oss cards describe a 21B and a 117B model, and Meta's Llama 4 card lists Scout at 109B and Maverick at 400B total.

The table shows one current model per family. Families ship new versions every few months, so read the names as examples and check the publisher's page for the newest. Parameters are the weights counted in billions, so 21B means 21 billion. Where a model is a mixture of experts, the table also gives its active share, a term that [lesson 3](https://mindket.com/learn/open-weight-models/choose-an-open-model) explains.

| Family | Published by | Model read on 6 Nov 2026 | Parameters | What its card calls it |
| --- | --- | --- | --- | --- |
| gpt-oss | OpenAI | gpt-oss-20b | 21B, 3.6B active | For lower latency, and local or specialized use cases |
| Qwen | Qwen | Qwen3.8-27B | 27B | A native vision-language model |
| DeepSeek | DeepSeek | DeepSeek-V4.1-Flash | 552B backbone | A multimodal mixture-of-experts model |
| Llama | Meta | Llama 4 Scout | 109B, 17B active | Multilingual text and image input |
| Gemma | Google DeepMind | Gemma 4 12B | 11.95B | Text, image and audio input |
| Mistral | Mistral | Mistral Large 4 | About 1T, 49B active | A natively multimodal mixture-of-experts model |

## Common mistakes

- Treating "open" in a headline as permission. The license on the card says what you may do, and lesson 2 shows how to read it.
- Calling a weights-only release open source in your own writing. Say open-weight and name the license.
- Downloading before checking size. In llama.cpp's table a 405B model is 1,625.1 GB before quantization.
- Skipping the access request on a gated model. Meta's Llama 4 page asks for a login and a form, and the files stay locked until the request is approved.
- Copying a model name from a tutorial. Names such as Qwen3.8 or Gemma 4 stop being current as new versions arrive, so read the publisher's page.

## Try it yourself

- Open the pages for gpt-oss-20b and Olmo 3.1 32B Think and list what each repository holds that the other does not.
- Find the license field on the Qwen3.8-27B and DeepSeek-V4.1-Flash cards and write both down.
- Run `ollama run gpt-oss:20b` if you have 16GB of memory free, after the Ollama tutorial.
- Pick one family above and write down the first three things you would check before using it at work.

## Next: open-model licenses

Every card carries a license field, and the six models above hold four different kinds. [Lesson 2](https://mindket.com/learn/open-weight-models/open-model-licenses) reads them line by line.

## Check yourself

1. What do the weights of a model contain?

   A. The text it was trained on
   B. The code that ran the training
   C. The numbers it learned in training
   D. Its license terms

   Answer: C. The weights are the numbers learned during training, stored in files such as the `.safetensors` ones.

2. Which of these makes a release more than open-weight?

   A. A license file in the repository
   B. A longer model card
   C. A download command for Ollama
   D. Released training code and named training datasets

   Answer: D. Open source adds the training code and details of the data to the weights, as the Olmo 3.1 card does by naming Dolma 3 and Dolci.

3. Where do you read the license a maker chose for a model?

   A. In the file sizes
   B. In the card's license field and the LICENSE file
   C. In the tokenizer files
   D. In the download count

   Answer: B. The card's settings block holds a license field, and the repository usually carries the full text in a LICENSE file.

4. Which files in a model repository are the weights?

   A. Files ending in .safetensors
   B. LICENSE and USAGE_POLICY
   C. tokenizer.json
   D. README.md

   Answer: A. On gpt-oss-20b the `.safetensors` files hold the weights. The others are the license, the tokenizer and the card.

## Frequently asked questions

### What is the difference between open weights and open source?

Open weights means the trained parameters are published for download. Open source adds the full training code and a description of the data, under terms that let you study, change and share the system. Some models called open source only meet the first half.

### Is Llama open source?

Meta calls it open source, but open-weight is the accurate word. The weights come under the Llama 4 Community License, which asks for a "Built with Llama" notice and a license from Meta above 700 million monthly users. The Open Source Initiative and others have contested the label because the license carries an acceptable use policy that bars some uses.

### Can I run an open-weight model on my own computer?

Yes, if it fits in memory. The gpt-oss-20b card says it runs within 16GB, while the larger gpt-oss-120b targets a single 80GB GPU. The [local hardware lesson](https://mindket.com/learn/local-ai/local-llm-hardware) shows how to size a model, and lesson 4 covers the quantization that shrinks it.

### Are open-weight models free to use in a business?

Often yes, under terms that differ by license. Apache 2.0 and MIT allow commercial use with a notice. Meta's license adds a condition above 700 million monthly users, and Mistral's limits hosting the model for others as a service. Lesson 2 compares them.

### What file format are the weights in?

Hugging Face lists safetensors as a recommended model format, and gpt-oss-20b's weights are .safetensors files. GGUF files hold the weights and extra metadata in one file for llama.cpp, Ollama and LM Studio, and lesson 4 covers them.

## Sources

- [GitHub: openai/gpt-oss README](https://github.com/openai/gpt-oss) (read 2026-11-06)
- [Hugging Face: openai/gpt-oss-20b model card](https://huggingface.co/openai/gpt-oss-20b) (read 2026-11-06)
- [Hugging Face API: openai/gpt-oss-20b (file list)](https://huggingface.co/api/models/openai/gpt-oss-20b) (read 2026-11-06)
- [Hugging Face: allenai/Olmo-3.1-32B-Think model card](https://huggingface.co/allenai/Olmo-3.1-32B-Think) (read 2026-11-06)
- [GitHub: Meta Llama 4 model card](https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md) (read 2026-11-06)
- [Hugging Face: meta-llama/Llama-4-Scout-17B-16E-Instruct model page (gated)](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) (read 2026-11-06)
- [Hugging Face Docs: Gated models](https://huggingface.co/docs/hub/models-gated) (read 2026-11-06)
- [Hugging Face: Qwen/Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B) (read 2026-11-06)
- [Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) (read 2026-11-06)
- [Hugging Face: google/gemma-4-12B-it model card](https://huggingface.co/google/gemma-4-12B-it) (read 2026-11-06)
- [Hugging Face: google/gemma-4-26B-A4B-it model card](https://huggingface.co/google/gemma-4-26B-A4B-it) (read 2026-11-06)
- [Hugging Face: mistralai/Mistral-Large-4-1T-A52B model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B) (read 2026-11-06)
- [Wikipedia: Open-source artificial intelligence](https://en.wikipedia.org/wiki/Open-source_artificial_intelligence) (read 2026-11-06)
- [Hugging Face Docs: Model Cards](https://huggingface.co/docs/hub/model-cards) (read 2026-11-06)
- [Hugging Face Docs: Licenses](https://huggingface.co/docs/hub/repositories-licenses) (read 2026-11-06)
- [Hugging Face Docs: GGUF](https://huggingface.co/docs/hub/gguf) (read 2026-11-06)
- [llama.cpp: quantize README](https://github.com/ggml-org/llama.cpp/blob/master/tools/quantize/README.md) (read 2026-11-06)

[Browse kits](https://mindket.com/kits)