# What are open-weight models?

> An open-weight model is an AI model whose trained weights anyone can download. It is not the same as open source, and the license decides what you may do.

Source: https://www.honeyhive.ai/learn/open-weight-models (updated 2026-11-16)

An open-weight model is an AI model whose trained parameters, its weights, are published so that anyone can download them. Stanford HAI [describes](https://hai.stanford.edu/ai-definitions/what-is-an-open-weight-model) it as a model whose core components are publicly released, which lets users run it on their own computers, study how it works and modify it. The Open Source Initiative (OSI) [puts it](https://opensource.org/ai/open-weights) more narrowly: "the final weights and biases of a trained neural network."

Open weights are not the same as open source, and "open" does not tell you what the license allows. Both points decide whether you can build on a model.

## Open weights versus open source

The OSI's [Open Source AI Definition](https://opensource.org/ai/open-source-ai-definition) (version 1.0) asks for the freedom to use, study, modify and share a system. It requires three things to be available in the preferred form for making changes:

- **Data information:** "sufficiently detailed information about the data used to train the system."
- **Code:** "the complete source code used to train and run the system."
- **Parameters:** the weights or other configuration settings.

Open weights cover only the third item. The OSI [says](https://opensource.org/ai/open-weights) open weights "fall short of this because they do not provide the underlying training process, code, or comprehensive data details," and lists training code, the training dataset and intermediate checkpoints as the usual gaps. It places open weights in between: "more transparent than purely proprietary AI," yet lacking "several key elements of Open Source AI."

| | Closed (API only) | Open weights | Open source AI (OSI) |
|---|---|---|---|
| Weights downloadable | No | Yes | Yes |
| Training code | No | Usually no | Required |
| Training data information | No | Usually limited | Required |
| Terms | Provider terms | The model's license | Terms granting use, study, modify and share |

Table summarizes the OSI's [open weights](https://opensource.org/ai/open-weights) and [Open Source AI Definition](https://opensource.org/ai/open-source-ai-definition) pages; "usually" marks the OSI's description of typical releases, not a rule every release follows.

## The license sets the terms

Having the file does not give you the right to use it however you like. The license is the contract, and the terms vary a lot. Hugging Face lets a model repository [declare any license](https://huggingface.co/docs/hub/en/repositories-licenses) in its card metadata, from `apache-2.0` and `mit` to `cc-by-nc-4.0` and model-specific ones such as `llama4` and `gemma`. For anything outside its list, the repo sets `license: other` and ships the license text.

Two examples show the range:

- **Apache 2.0.** OpenAI's gpt-oss models list [`apache-2.0`](https://huggingface.co/openai/gpt-oss-120b) as their license. The [license text](https://www.apache.org/licenses/LICENSE-2.0) grants a perpetual, worldwide, royalty-free copyright license that covers derivative works and distribution, and ends your patent license if you file patent litigation claiming the work infringes a patent.
- **A custom community license.** The [Llama 4 Community License](https://dev.meta.ai/llama/llama4/license/) (effective April 5, 2025) requires a license request to Meta if your products had more than 700 million monthly active users in the month before the release date, and Meta may deny it. If you distribute the model or a product that contains it, you must display "Built with Llama"; a model trained or fine-tuned with it that you distribute must start its name with "Llama"; and all use must follow Meta's acceptable use policy.

Read the actual license for the exact model and version, since terms differ between releases.

## What open weights let you do

- **Run it where you choose.** Stanford HAI [notes](https://hai.stanford.edu/ai-definitions/what-is-an-open-weight-model) that users can run the model on their own computers, or you can pick a host, so prompts need not go to the model's maker.
- **Fine-tune it.** The OSI [notes](https://opensource.org/ai/open-weights) that shared parameters let others fine-tune, adapt or deploy the model. See [prompt engineering vs fine-tuning](/learn/prompt-engineering-vs-fine-tuning) before you commit to training.
- **Study it.** You can inspect the architecture and weights, though without the training data you cannot fully explain why it behaves as it does.

The tradeoffs are practical. You carry the serving cost, and each release has its own required formats: gpt-oss, for example, [should only be used](https://huggingface.co/openai/gpt-oss-120b) with its harmony response format. Nothing runs application-level checks for you, so plan your own [guardrails](/learn/ai-guardrails).

## Why the term is debated

Policy reports use the same idea under another name. The US NTIA's [report](https://www.ntia.gov/issues/artificial-intelligence/open-model-weights-report) (July 30, 2024) covers models with "widely available" weights, meaning users can download them. It said such models can "decentralize AI market control from a few large AI developers" and also could "engender harms and risks" to areas such as safety and privacy, and concluded that "current evidence is not sufficient to definitively determine" whether restrictions are warranted.

## Choosing one for an application

Questions to answer before you pick a model:

1. Does the license allow your use, including commercial use, redistribution and any naming or attribution rule?
2. What do you need beyond the weights: training data details, code, or only the ability to run it?
3. Can you serve it, or will you pay a host? See [LLM cost optimization](/learn/llm-cost-optimization).
4. How will you compare it with a hosted model on your own task? Use your own [evaluation](/learn/llm-evaluation), since public scores may not match your data.

## Tracing open-weight models

Whichever model you choose, record its calls so you can compare it with the alternatives. HoneyHive's [multi-provider tracing](https://docs.honeyhive.ai/v2/tracing/multi-provider.md) attaches each provider's instrumentor to one tracer, so spans from several providers land in the same trace. The page covers hosted providers, so check the [tracing docs](https://docs.honeyhive.ai/v2/tracing/introduction.md) for how to instrument a model you serve yourself. More background in [LLM observability](/learn/llm-observability).

## FAQ

### What is an open-weight model?

A model whose trained parameters are published so anyone can download them. The Open Source Initiative [defines](https://opensource.org/ai/open-weights) open weights as the final weights and biases of a trained neural network.

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

Open source AI, as the Open Source Initiative [defines it](https://opensource.org/ai/open-source-ai-definition), also requires the complete training code and sufficiently detailed information about the training data. Open weights give you the parameters and usually not the rest.

### Can I use open-weight models commercially?

It depends on the license. Apache 2.0 [grants](https://www.apache.org/licenses/LICENSE-2.0) a royalty-free copyright license that covers derivative works and distribution. Other licenses add conditions, such as the Llama 4 Community License's [700 million monthly active user threshold](https://dev.meta.ai/llama/llama4/license/).

### Are open-weight models free?

The download can be free, but running the model is not. You supply the hardware or pay a host, and you maintain the deployment yourself.

### Can I fine-tune an open-weight model?

Usually yes, if the license allows it. The OSI [notes](https://opensource.org/ai/open-weights) that parameters shared under an OSI Approved License let others fine-tune, adapt or deploy the model.
