Open-Source LLMs vs Closed Models: A 2026 Comparison

Two years ago the choice was easy to explain. Closed models from the large labs were clearly better, and open models were what you used when you could not send data out or could not afford the bill. That gap has narrowed. Stanford's 2025 AI Index reported that open-weight models cut their performance difference with closed models from 8% to 1.7% on some benchmarks in a single year (Stanford HAI: 2025 AI Index).

So the decision now turns on other things: who controls the model, what the licence allows, where the data goes and how the cost behaves. This comparison works through each.

What "open" means for a language model

Most models called open source are, more precisely, open-weight models. You can download the trained parameters and run them yourself. You usually cannot see the training data or rebuild the model from scratch.

The Open Source Initiative has defined the stricter term. Its Open Source AI Definition requires the freedoms to use, study, modify and share a system for any purpose. It also requires access to what is needed to modify it properly: detailed information about the training data, the complete code used to train and run it, and the parameters, all under approved terms (OSI: Open Source AI Definition). Few popular models meet all of that.

For practical decisions, the weights and the licence matter most. Open weights decide whether you can run and change the model. The licence decides what you may do with it.

The licence is the product

Open-weight models come under very different licences, and the differences are commercial.

The details, read from the licences themselves:

• gpt-oss is released under Apache 2.0, and the larger model is described as fitting on a single 80GB GPU, the smaller one within 16GB of memory (Hugging Face: gpt-oss-20b).
• Qwen2.5-7B-Instruct is Apache 2.0, with a context length of 131,072 tokens (Hugging Face: Qwen2.5-7B-Instruct). Its 3B sibling is published under a licence named qwen-research (Hugging Face: Qwen2.5-3B-Instruct). One family, two sets of terms. Read the card of the exact model you deploy.
• Llama 3.2 requires anyone distributing it, or a product containing it, to include the agreement and display "Built with Llama". Use must follow Meta's acceptable use policy. A company whose products had more than 700 million monthly active users when the model was released must ask Meta for a licence (Meta: Llama 3.2 Community License).
• Gemma may be used, modified and distributed, but you must pass its use restrictions on to recipients. Google reserves the right to restrict usage it believes breaks the terms, and claims no rights in the outputs (Google: Gemma Terms of Use).

None of these is a reason to avoid a model. Each is a reason to have someone read the licence before a product depends on it.

What a closed model gives you

A closed model is used through the vendor's API. You never hold the weights.

• The strongest models, usually, at the frontier of capability.
• No infrastructure. Scaling, hardware and serving are the vendor's problem.
• Managed features such as tool calling, structured output and long contexts, maintained for you.

What you accept in return:

• Your data leaves your network. Read the vendor's data terms. OpenAI, for example, states that API data has not been used to train its models since 1 March 2023 unless you opt in. Abuse-monitoring logs, which can contain prompts and responses, are kept for up to 30 days by default. Eligible customers can apply for zero data retention (OpenAI: data controls).
• Models are retired on the vendor's schedule. OpenAI's policy gives at least six months' notice for generally available models and less for preview models (OpenAI: deprecations). A prompt tuned for one model must be re-checked on its replacement.
• Pricing per token, which you do not control.

What open weights give you

• Control over where it runs: your cloud, your data centre or a laptop. Sensitive data never has to leave.
• A fixed version. The model you tested is the model you run until you choose to change it.
• Full customisation, including fine-tuning. See LoRA and QLoRA.
• Small models for small jobs. A 3-billion-parameter model runs on an ordinary laptop CPU, as in the experiments in few-shot vs zero-shot prompting.

And the costs:

• Operations are yours: GPUs, serving, scaling, monitoring. See vLLM and model serving.
• Safety is yours. A hosted API comes with the vendor's filters. A downloaded model comes with whatever you add.
• The largest models need serious hardware, and capability still tends to arrive first in closed models.

How the cost behaves

The two are priced on different shapes. An API charges per token, so cost tracks usage exactly and is near zero when traffic is. A self-hosted model costs the price of its hardware for every hour it is running, busy or idle, plus the engineering time to run it.

That makes utilisation the deciding number. A GPU kept busy by steady, heavy traffic can produce tokens more cheaply than an API. The same GPU serving a few hundred requests a day is mostly an idle bill. Compare the two with your own traffic: tokens per month, how evenly they arrive, and what an engineer's time costs. See cost optimization for LLM applications.

Choosing

Many teams end up with both: a closed model for the hardest requests, an open one for high-volume or sensitive ones. That is easier than it sounds, because popular local and self-hosted runners offer OpenAI-compatible APIs, so switching is often a change of address and model name. See the Ollama tutorial.

Whatever the shortlist, decide with your own test set, not a leaderboard. A model that ranks higher in general can still lose on your documents and your questions. See evaluating LLM outputs.

Common mistakes

• Treating "open" as "open source" and assuming no restrictions apply.
• Reading the family's licence, not the model's. Sizes in one family can differ.
• Choosing on benchmarks alone.
• Counting GPU price but not engineering time when comparing costs.
• Ignoring a closed vendor's data terms until a security review asks.
• Building on a preview model that can be retired at short notice.

Working with open models through Hugging Face, and deploying models to the cloud, are part of the Generative AI course.

Sources

• The Open Source AI Definition 1.0 (Open Source Initiative) — Open Source Initiative
• Llama 3.2 Community License Agreement (GitHub, Meta) — Meta
• Gemma Terms of Use (Google AI for Developers) — Google
• gpt-oss-20b model card (Hugging Face) — OpenAI
• Qwen2.5-7B-Instruct model card (Hugging Face) — Qwen team, Alibaba Cloud
• Qwen2.5-3B-Instruct model card (Hugging Face) — Qwen team, Alibaba Cloud
• Data controls in the OpenAI platform (OpenAI API documentation) — OpenAI
• Deprecations (OpenAI API documentation) — OpenAI
• The 2025 AI Index Report (Stanford HAI) — Stanford HAI