An open-weight AI model is one whose trained numerical parameters — the billions of numbers that encode what the model has learned — are made publicly available for download. Anyone can take those weights and run the model on their own hardware, adapt it to a new task, or study how it behaves. Closed models such as GPT-4 or Gemini keep those parameters private; users interact with them only through a pay-per-query API managed by the provider.

The distinction matters because the weights are the model. Releasing them shifts control from the provider to whoever downloads them.

Open-weight is not the same as open-source

The terms are often used interchangeably, but they mean different things. A fully open-source AI model would release not just the weights but also the training code, the training data, and full documentation — everything needed to reproduce the model from scratch.

Most publicly released models are open-weight only. Meta’s Llama 4, for example, releases weights but not the training dataset. Mistral’s models publish weights and some architecture details, but training data remains proprietary. True open-source AI — where all components are transparent and reproducible — is rare; the Open Source Initiative formally defined the standard in 2024, and most popular “open” models do not yet meet it.

How closed models work differently

With a closed model, you send a prompt to the provider’s servers, the model generates a response, and you receive it back. You never see the model’s parameters, you cannot modify its behavior directly, and you pay for each query. The provider controls updates, pricing, and availability.

With an open-weight model, you download the parameters to your own machine — or a cloud server you control — and run inference locally. Your data stays on your infrastructure, you are not subject to per-token pricing at scale, and you can fine-tune the model on your own datasets.

Why companies release open-weight models

Different players have different motivations. Meta releases Llama because wider adoption establishes it as an industry standard, and community researchers improve the model in ways Meta’s own team would not find quickly. Mistral, a Paris-based startup, uses open releases to build credibility against larger competitors. Chinese labs — including Zhipu AI (GLM), Alibaba (Qwen), and DeepSeek — have released open-weight models partly to build developer ecosystems independent of US API providers, and partly because their efficient training methods give them a cost advantage in inference.

By mid-2026, Chinese open-weight models account for a majority of tokens processed on neutral AI model-routing platforms.

What you can do with an open-weight model

Run it privately. A hospital, a law firm, or a government ministry can deploy a capable language model without sending sensitive documents to a third-party server.

Cut costs at scale. At high query volumes, running an open-weight model on owned or rented GPU infrastructure typically costs significantly less than equivalent commercial API calls — the savings grow with usage.

Fine-tune for your domain. Because you hold the weights, you can continue training on proprietary data to specialize the model for a specific task — legal drafting, customer support, or local-language applications.

Try it without a subscription. Ollama lets you download and run models such as Llama or Mistral on a modern laptop in minutes, with no API key or account required.

The risks and trade-offs

Open-weight release also has genuine downsides. Because anyone can download and modify the model, the safety constraints baked in at training time can be removed. Researchers have demonstrated that fine-tuning with a small number of adversarial examples is enough to strip most safety guardrails. Harmful uses of the underlying model become harder to trace or stop.

There are also practical costs. Running capable models requires significant GPU memory; the largest open-weight models demand data-center hardware. Hosting, monitoring, and keeping a model up to date requires engineering effort that a simple API subscription avoids.

On raw capability, closed frontier models still hold an edge on the most complex multi-step reasoning tasks, though that gap has narrowed significantly since 2024.

In the news

The release of GLM-5.2 by Z.ai — an open-weight model that scored competitively with closed frontier models at roughly one-sixth the API cost — illustrates how quickly open-weight models are closing the performance gap.

FAQ

Can I use an open-weight model commercially?
Usually yes, but check the license. Llama 4 carries a Meta custom license that restricts use above a certain monthly user threshold. Mistral and GLM-5.2 use Apache 2.0 or MIT licenses, which allow commercial use freely.

Do I need expensive hardware?
Smaller models (7B–13B parameters) run on a modern laptop with 16–32 GB of RAM. Larger models (70B+) typically need a dedicated GPU with substantial VRAM. Cloud providers offer GPU instances by the hour if you do not want to buy hardware.

Is open-weight the same as free?
The weights are free to download. Running them still costs electricity and compute — either your own hardware or rented cloud resources.

Are open-weight models safe to use?
Reputable models from established labs include safety training, but this can be removed or weakened by downstream fine-tuning. For sensitive applications, apply your own content filtering and evaluate the model’s behavior thoroughly for your use case before deployment.