What is Open Weights? — AI Glossary

What it is: An open-weight AI model is one where the trained model files are publicly downloadable. Anyone can run the model locally, fine-tune it, or integrate it without paying per-token API fees. Llama, Mistral, DeepSeek, and Gemma are all open-weight.
Who it is for: Developers, researchers, and businesses that want AI without ongoing API costs or vendor lock-in.
Best if: You want to run AI on your own hardware (for privacy, cost, or customization), or you’re comparing the open-source AI ecosystem to closed APIs.
Skip if: You’re a casual user happy with ChatGPT, Claude, or Gemini consumer apps. Want one practical AI workflow every morning? Subscribe to our free daily newsletter.

What are open weights?

An open-weight AI model is one whose trained model files (the weights — the actual numerical parameters that make the model work) are made publicly available for download. Anyone can take those files, run the model on their own hardware, fine-tune it, embed it in a product, or modify it.

Open weights is different from open source software. Open source typically requires the source code, training data, and processes to be available. Open weights releases the trained model but often without the original training data or full training code. You can use and adapt the model, but you can’t necessarily reproduce its training from scratch.

Major open-weight model families in 2026: Llama (Meta), Mistral, DeepSeek, Gemma (Google), Phi (Microsoft), Qwen (Alibaba). Most are licensed under permissive but not fully open licenses — some restrict commercial use at huge scale, for example.

Why do open weights matter?

Open weights changed the economics of AI. Before Meta released Llama 1 in early 2023, capable AI was effectively only available through paid APIs to OpenAI, Anthropic, and Google. After Llama, anyone could download a comparably capable model and run it themselves — no API costs, no rate limits, no data leaving your infrastructure.

That unlocked entire ecosystems: thousands of fine-tuned model variants on Hugging Face, local-AI tooling (Ollama, llama.cpp, LM Studio), enterprise self-hosting, and a generation of startups building on open weights rather than paying per-token. It also shifted the competitive dynamics — if open-weight models are good enough for most use cases, closed-API companies can’t charge premium prices for routine work.

How do open weights compare to closed weights?

The two are different business models:

• Open weights — weights downloadable, free or low-cost to run yourself, no vendor lock-in, requires your own infrastructure. Examples: Llama, Mistral, DeepSeek.
• Closed weights — weights stay with the company, accessed only via API, pay per token, no infrastructure to manage. Examples: GPT, Claude, Gemini Pro.

Many companies use both. Closed APIs for the most capability-demanding work (frontier reasoning, complex multi-step tasks), open weights for routine work where cost and privacy matter more than peak capability. In 2026 the gap between open and closed has narrowed dramatically — DeepSeek and Llama can match GPT-4-class capability for most tasks, though the absolute frontier still belongs to the closed labs.

Related terms

• Closed Weights
• Llama
• Mistral
• DeepSeek
• Hugging Face

Learn more on Beginners in AI

• Ollama — Run Open-Weight AI Locally
• AI Glossary

Sources and further reading

• Hugging Face — primary open-weight hub
• Meta Llama — flagship open-weight family

Last reviewed: May 2026. AI terminology evolves quickly — verify specifics on the official source pages above.

You may also like

• Ollama — Run Open-Weight AI Locally
• AI Glossary
• Closed Weights
• Llama
• Mistral
• DeepSeek