Open-weight vs closed AI models

Decision table

What "open weight" actually means

A model's weights are the learned numbers that make it work. An open-weight release lets you download them and run the model yourself. That is narrower than open source.

The Open Source Initiative's Open Source AI Definition requires the freedom to use, study, modify and share a system "for any purpose," plus "sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system." Most open-weight releases publish weights and code but not that level of training data detail.

Licenses vary more than most buyers expect:

• MIT: DeepSeek's V4.1 Flash model card states the repository and weights are MIT-licensed, one of the most permissive terms.
• Apache 2.0: OpenAI's gpt-oss-120b model card and Mistral's Small 3 announcement both use it.
• Custom community licenses: Meta's Llama 3 license requires a separate license from Meta for any company whose products had more than 700 million monthly active users at release.
• Use-restricted terms: Google's Gemma terms bind users and anyone they redistribute to a prohibited use policy.

Meta's Llama 3 license requires companies with more than 700 million monthly active users to request a separate license, according to the license text checked 17 October 2026.

Steps to decide

1. Write down the hard constraints first. Data residency, air-gapped deployment and audit duties often settle the question before quality does.
2. Read the license, not the headline. Check commercial use, user caps, redistribution duties and acceptable use clauses.
3. Size the hardware. Total parameters set memory needs; active parameters drive speed in mixture-of-experts models. gpt-oss-120b has 117B parameters with 5.1B active and fits one 80GB GPU. DeepSeek V4.1 Flash has 552B backbone parameters with 8B active in prefill and 16B in decode.
4. Price the hosted version first. Many open models are also sold as APIs. If a hosted API meets your data terms, it is usually cheaper than owning idle GPUs.
5. Run the same test set on both. Use 30 to 50 of your real tasks, grade them blind and compare pass rates.
6. Count the people cost. Self-hosting needs someone to patch, monitor, scale and update the model.
7. Plan your exit. Closed models retire on vendor schedules; open weights you host stay fixed until you change them.

Worked example: a law firm summarizing contracts

A 40-lawyer firm wants contract summaries and clause extraction. Client confidentiality rules bar sending documents to third parties without contractual safeguards.

• Option A, closed API: Claude Sonnet 5.5 at USD 2 input and USD 10 output per 1M tokens, as listed 17 October 2026, through a cloud provider the firm already contracts with. No hardware, strong long-document handling, and the firm relies on the provider's enterprise data terms.
• Option B, open weights on premises: gpt-oss-120b on one 80GB GPU server. Documents never leave the building, the license is Apache 2.0, and the firm controls updates. It needs an IT owner and a quality check against Option A.
• Option C, open weights via a hosted API: cheapest per token, but data still leaves the firm, so it fails the confidentiality constraint unless terms are negotiated.

Our verdict for this firm: Option B if the IT team can run it and the test set shows acceptable summaries; otherwise Option A under enterprise terms. Option C does not meet the stated constraint.

Policy and regulation

The US National Telecommunications and Information Administration's report on open model weights (July 2024) did not recommend restricting widely available weights. It recommended that the government "actively monitor a portfolio of risks" and be ready to act if they rise.

In the EU, Article 53 of the AI Act exempts providers of general-purpose models released under a free and open-source license, with weights and architecture public, from two documentation duties. The exemption does not apply to models with systemic risk. The European Commission's AI Act page says obligations for general-purpose models applied from 2 August 2025. If you deploy either kind of model in the EU, your own deployer duties still apply.

Common mistakes

• Calling every downloadable model "open source." Most do not meet the OSI definition; say open weight.
• Assuming free weights mean a free system. GPUs, power, engineers and monitoring cost money every month.
• Skipping the license's user cap or use policy. These can block a product at scale.
• Comparing a self-hosted quantized model to a full-precision API. Quantization can change quality; test the exact build you will run.
• Ignoring updates. An open model you never update will fall behind closed models that vendors replace on their own schedules.

How we researched this

We read the model cards and licenses for DeepSeek V4.1 Flash, gpt-oss-120b, Mistral Small 3, Llama 3 and Gemma; the OSI definition; NTIA's report; the AI Act text and Commission page; and vendor pricing pages on 17 October 2026. API prices match our model directory and the DeepSeek V4.1 Flash profile. For data handling questions, use our AI tool privacy checklist, and for engineering roles that run these models, see the machine learning engineer career guide.

Frequently asked questions

What is the difference between open-weight and open-source AI? Open weight means you can download the trained parameters. Open source, under the OSI definition, also requires detailed training data information and freedom to use, study, modify and share for any purpose.

Can I use DeepSeek models commercially? Yes, under the MIT License stated on DeepSeek's Hugging Face model cards for its V4 family. Check your own data and export rules separately.

Is self-hosting an open model cheaper than an API? Only at steady, high volume or when data rules require it. Hosted APIs such as DeepSeek's list off-peak input at USD 0.15 per 1M tokens, which is hard to beat with idle hardware.

What hardware do I need to run gpt-oss-120b? One 80GB GPU such as an NVIDIA H100, according to OpenAI's model card. The smaller gpt-oss-20b runs within 16GB of memory.

Are open-weight models less safe? They carry different risks, since safeguards can be removed once weights are public. NTIA's 2024 report recommended monitoring these risks rather than restricting release.

Does the EU AI Act treat open-source models differently? Yes, Article 53 exempts qualifying open-source general-purpose models from two documentation duties, except models with systemic risk.

Is Llama open source? Not by the OSI definition. Meta's Llama 3 license adds a 700 million monthly user threshold and an acceptable use policy.