Open-weight models for business: what they are and when to use one

An open-weight model is an AI model whose trained weights, the learned parameters that turn an input into an output, are published for anyone to download and run. OpenAI's gpt-oss, Meta's Llama, Google's Gemma and DeepSeek V4 are all open-weight models. A closed model, by contrast, is only available through its maker's app or API. For a business, the difference is control: an open-weight model can run on your own servers or inside your own cloud account, so prompts and customer data never go to the model's maker, and it can be fine-tuned on your own data. This answer explains what open weight means, how it differs from open source, what it costs to run one and when a brand should. It is from Brand Media Works, an AI consulting and marketing data engineering firm for ecommerce brands.

What is an open-weight model?

The Open Source Initiative defines AI weights as "the set of learned parameters that overlay the model architecture to produce an output from a given input" (read 4 November 2026). Training a model sets those parameters. Publishing them lets anyone run the finished model without retraining it.

OpenAI, for example, released gpt-oss-120b and gpt-oss-20b in August 2025 as open-weight models, with the weights free to download from Hugging Face. OpenAI says gpt-oss-120b reaches near-parity with its o4-mini model on core reasoning benchmarks while running on a single 80 GB GPU, and gpt-oss-20b runs with 16 GB of memory (read 4 November 2026). DeepSeek released V4 in April 2026 with its weights published (read 25 October 2026).

When should a brand run an open-weight model, and what does it cost?

There are three ways to run one, at published prices:

• On your own hardware. Ollama is free and does not meter models run locally, and its Pro plan for cloud models is $20 a month (read 15 October 2026). This suits testing more than production.
• Through a hosted API. Together AI lists gpt-oss-120B at $0.15 per million input tokens and $0.60 per million output tokens, and Gemma 4 31B at $0.39 and $0.97 (read 15 October 2026). This is the cheapest way to run one at scale, but prompts still go to a third party.
• Privately, in your own cloud. Amazon Bedrock's Custom Model Import charges $0.05718 per Custom Model Unit per minute, Llama 3.1 8B needs 2 units, and an idle model scales to zero after 5 minutes (read 15 October 2026). That is about $6.86 an hour while the model is active (our arithmetic). Together's dedicated single-tenant endpoints cost $5.49 per H100 GPU per hour.

Fine-tuning is affordable too. Together AI lists LoRA fine-tuning of Gemma 4 26B A4B at $1.05 per million training tokens (read 31 October 2026).

For most brands, a closed model through a business plan or a cloud provider is the simpler start. An open-weight model earns its setup cost when customer records or other sensitive data should never leave the brand's own account, when volume is high enough that per-token API fees add up, or when the brand wants a model fine-tuned on its own products and voice. It also helps with data residency. DeepSeek's privacy policy says it stores the personal data it collects in China (read 25 October 2026), but DeepSeek's open weights can run in any region the brand chooses. Private LLMs on open-weight models describes how Brand Media Works builds these inside a brand's cloud account, what a private LLM costs to run sets out the running costs, and fine-tuning vs RAG covers when to fine-tune.

Are open-weight models free?

The weights are free to download, but using them is not unconditional. Each model comes with a license (read 4 November 2026):

Running the model also costs money: hardware, cloud compute or per-token API fees. "Free" describes the license, not the bill.

What is the difference between open-weight, open-source and closed models?

• Closed model: the weights are private. You use the model only through its maker's app or API, under its terms and data policies.
• Open-weight model: the weights are published, so you can run and fine-tune the model yourself, under the license's conditions.
• Open source model: under the Open Source Initiative's definition, open source AI must grant the freedoms to use, study, modify and share the system. It must also include the complete code used to train and run it, plus enough information about the training data for a skilled person to build a substantially equivalent system.

Most models people call "open source" are open weight. They publish the weights but not the full training data and code.

What is the best open-weight model?

There is no single best one. The right model depends on four things:

1. License: whether its terms fit how the brand will use and distribute it.
2. Size: whether it fits the hardware or budget. gpt-oss-20b runs in 16 GB of memory, while DeepSeek V4-Pro has 1.6 trillion total parameters (read 25 October 2026).
3. Hosting: whether it is available where the brand wants to run it, such as Bedrock or a dedicated endpoint.
4. Results on your own tasks: a small test set of real product questions or support tickets says more than public benchmarks.

Book a discovery call

Frequently asked questions

Is an open-weight model safer for customer data?

It can be, if it runs in the brand's own account, because prompts never go to the model's maker. Run through a third-party API, it carries the same data questions as any hosted model. The safety comes from where it runs, not from the weights being open.

Is Llama open source?

Llama is open weight. Its weights are published under Meta's community license, which adds conditions such as the 700 million monthly user threshold and the "Built with Llama" notice. It does not publish the full training data and code that the Open Source Initiative's definition requires.

Can open-weight models be misused?

Yes. OpenAI notes that once an open-weight model is released, bad actors may be able to fine-tune it for harmful purposes. That is why model makers test before release and attach acceptable use terms.