Open-weight AI models are yours to run and to maintain.

“Should we use an open-source model?” A client asks it as if there were one answer. The phrase covers three different things: a model you can download and run yourself, a license that sets terms on what you do with it, and, rarely, a full release of everything that went into building it. Most of what gets called open-source Artificial Intelligence (AI) is the first one, and the useful question is what that gets you.

What does “open-weight” hand you?

The weights: the learned numbers that make up the finished model. With them, you can run it on a machine you control. OpenAI's gpt-oss models ship under the Apache 2.0 license, with a minimal usage policy¹, and Mistral said of its first 7B model that under Apache 2.0 it can be used without restrictions⁴. Meta's Llama is different. Its community license asks anyone whose products pass 700 million monthly active users in the preceding calendar month⁵ to request a license from Meta, and requires distributors to prominently display “Built with Llama”⁵. Same word, three sets of terms. Read the license before you read the benchmarks.

A weights release is not a full recipe. It gives you the finished model, and does not by itself give you the data it learned from. That matters less for running and tuning one, and more if a client asks you to explain where its knowledge came from.

Have open models caught up with the closed ones?

On the one public yardstick Stanford tracks, nearly. In January 2024 the leading closed model outperformed the top open-weight model by 8.04%⁶ on the Chatbot Arena leaderboard. By February 2025, this gap had narrowed to 1.70%⁶. A leaderboard is a general-knowledge contest, though, and your client's work is narrower than that. The gap that counts is the one on your own examples, which is an evals¹³ question.

What do you gain by running it yourself?

Control over where the data goes. Hugging Face, introducing gpt-oss, notes that many use cases rely on private and/or local deployments¹, and an open-weight model is how you do that: the prompts never leave your environment. The hardware bar is lower than many assume. The smaller gpt-oss model runs within 16GB of memory¹, and the larger one fits on a single H100 GPU¹. That bar is about size, though: owning the hardware now costs more and can take longer to get³, because AI data centers are buying up memory chips and GPUs². Price it at today's rates. You can also change the model itself. OpenAI describes gpt-oss as letting you fully customize models to your specific use case through parameter fine-tuning⁷, the heavier option from our fine-tune, retrieve or just ask better¹⁴ piece.

What does renting cost you that isn't on the invoice?

Control over the terms. With a closed model, the provider decides how it behaves, what it may be used for and how long each version stays available, and your pipeline changes when theirs does. Anthropic gives at least 60 days' notice⁹ before it retires a model, and after that requests to retired models fail⁹; OpenAI retired GPT-4o and older models¹⁰ from ChatGPT on February 13, 2026. Behavior moves too: in late October 2026, Anthropic added two sentences to Claude Code's prompt for Opus 5.5 to steer it away from recreating someone else's specific game¹¹, citing its usage policy on intellectual property. A small business that builds digital board-game modules for clients who own the games then reported that sessions that performed this work for months now refuse to continue it¹². Each change had a stated reason, and each still landed in someone's working pipeline. The open-weight model you downloaded stays as it is until you change it.

And what lands on your plate?

Everything a vendor's API quietly did for you. Somebody has to host the model, size the machines, apply updates, and watch it under load. Somebody has to build the layers that keep its output on brand and safe, because a raw model arrives without your client's rules. The guardrails¹⁵ are yours to build. Closed APIs bundle that work into the price per token¹⁶; with open weights you pay for it in people and servers instead. It can still come out cheaper at steady, high volume. It rarely comes out free.

What's Real?

• •“Open” is a spectrum, and the license is where it shows. Apache 2.0 comes with a minimal usage policy¹; Llama's terms carry a user-count threshold and an attribution rule. Treat the license as part of the product.
• •The quality gap is small on one public measure. Stanford's figure fell from 8.04% to 1.70%⁶ in about a year. That says nothing yet about your client's task.
• •Privacy and customization are the real draws. Local deployment keeps prompts in-house, and weights you hold can be tuned. Those are the reasons that survive scrutiny; “it's free” does not.
• •The EU AI Act does not give open models a free pass. Its exemption covers providers whose model is released under a free and open-source⁸ license with public weights, and it covers only some of the provider duties. It shall not apply to general-purpose AI models with systemic risks⁸. If you deploy one inside a product, ask your lawyer which duties are now yours.

The “so what” — for anyone serving tech clients

• •Ask what “open” means in the pitch. Weights only, or code and data information too? Under which license, and does the client's scale trip a threshold? Get the answers in writing before the architecture is drawn.
• •Price the whole bill. Put hosting, an owner for updates and the safety layer on the same sheet as the closed API's per-token cost, then compare at the client's real volume.
• •Run a bake-off on the client's own examples, closed and open side by side. The right-sized model¹⁷ is the cheapest one that clears the bar, and sometimes that is one you host.
• •If you rent, plan for the provider's changes. Pin a dated model version where the API allows it, keep that set of client examples to re-run whenever the provider announces a change, and know which model you would switch to.