# Open-source, general-purpose and specialized AI models in oil and gas: what is the difference?

> Oil and gas operators choose among three kinds of AI model: general-purpose proprietary models used through a provider's service, open-weight or open-source models whose weights can be downloaded and run on the operator's own servers, and specialized models adapted or trained for a domain. None removes the need to test the model on the operator's own engineering tasks.

- Label: Fact
- Type: Explainer
- Published: 2026-11-15T00:16:49+09:00
- By: Jun Zhong
- Researched and drafted by: Beacon's AI newsroom
- Checked by: Jun Zhong on 2026-11-15
- Canonical: https://beacon.winda.ai/library/open-source-vs-specialized-models/

Two benchmarks of language models on oil and gas subjects, published months apart in 2026, point in opposite directions. On 505 petroleum geoscience questions, the top models scored above 97% and several lower-cost open-weight models passed 90% [14]. On 1,200 petroleum engineering questions covering production, reservoir and drilling work, the best overall scores were 72% to 74% [13]. The test moves the answer as much as the model does.

## How do the three kinds of model compare?

A domain-specialized model is usually one of the other two, adapted further on domain text or task data [11].

| Dimension | General-purpose proprietary model | Open-weight or open-source model | Domain-specialized model |
|---|---|---|---|
| Who holds the weights | The developer; users send inputs and get outputs through a web interface or an application programming interface (API) [3] | Anyone who downloads them; the developer cannot take them back [3] | Whoever did the adapting, on top of the base model's terms [8] |
| Licence | The provider's service terms | From an OSI-approved licence to a community licence with use limits [1] [8] | Inherits the base model's licence, plus the adapter's own terms [8] |
| Where it can run | The provider's or a cloud's infrastructure, unless the vendor offers on-premises operation [5] | The operator's own hardware; comments to NTIA cited single graphics processors and even ordinary processors [3] | Wherever its base model can run |
| Who controls updates | The provider; versions can change behaviour within months [12] | The operator, who decides when to swap weights [3] | The party that retrains it |
| Evaluation burden | Public benchmarks plus the operator's own tests [4] | The same, and fewer built-in safeguards to rely on [3] | The same, plus checking the domain data it learned from [4] |

## What does "open source" mean for an AI model?

The Open Source Initiative (OSI) set out what counts as Open Source AI in version 1.0 of its definition. A system qualifies only if its terms let anyone use it for any purpose without asking, study how it works, modify it and share it [1]. Exercising those freedoms needs three things under OSI-approved terms: detailed information about the training data, the full source code for training and for running the model, and the parameters, meaning the weights [1].

Open weights are only the last of the three. OSI describes them as the learned values a network ends up with after training, and notes that releasing them without the training code and data details leaves others unable to reproduce or audit how the model was made [2].

Licences on downloadable models vary widely. The Apache License 2.0 grants a perpetual, royalty-free copyright and patent licence, with the patent licence ending for anyone who sues claiming the work infringes a patent [7]. The Llama 3.1 Community License, by contrast, requires a separate licence from Meta for licensees whose products had more than 700 million monthly active users on the release date, and requires "Built with Llama" to be displayed on products that use it [8]. Its acceptable use policy forbids using the model in activities that risk death or bodily harm, naming the operation of critical infrastructure and heavy machinery [9].

**Analysis:**
For an oil and gas operator the use clauses matter more than the user threshold. Whether drafting inspection reports counts as use related to operating critical infrastructure is a legal reading, and it should be settled and on file before the model goes near a plant [9].

## When does a general-purpose proprietary model fit?

When the data is not sensitive and the operator can accept changes on the provider's schedule. A 2023 study of one provider's March and June model versions found that accuracy at identifying prime numbers fell from 84% to 51% between them, and concluded that the "same" service can change substantially in a short time [12].

Data is the other limit. Joint guidance led by the US Cybersecurity and Infrastructure Security Agency (CISA), December 2025, advises not sharing sensitive data with AI models hosted by outside parties such as public cloud, and asking vendors whether they can run on premises or without continuous internet access [5]. It puts engineering configuration data, such as network diagrams and logic diagrams, first among the operational technology (OT) data to protect [5]. See [OT cybersecurity](/library/ot-cybersecurity/) for the wider rules.

## When does an open-weight or open-source model fit?

When control of the data and the version matters most. The US National Telecommunications and Information Administration (NTIA) noted in July 2024 that organizations running their own models gain more control over sensitive data used in fine-tuning than when they reach a model through an API [3]. Downloaded weights can be run locally, and NTIA notes that the developer then has no view of when or how they are used [3], which also means a copy can sit on a network with no outside link.

The trade-off is that safety work moves to the user. NTIA found significantly fewer model-based safeguards for open-weight models overall, and found that fine-tuning can strip the safeguards a developer added [3].

In the European Union, the AI Act's obligations for general-purpose AI models have applied since 2 August 2025 [6]. Providers releasing a model under a free and open-source licence, with its weights, architecture and usage information public, are exempt from the documentation duties, but not from the copyright policy or the training-content summary, and not at all if the model carries systemic risk [6].

## When does a domain-specialized model fit?

When general models miss the vocabulary or reasoning of the domain. A 2020 study across four domains found that a second round of pretraining on in-domain text improved results, with or without much labelled data [11]. Adaptation has become cheaper: Low-Rank Adaptation (LoRA) froze the base weights and, against full fine-tuning of a 175-billion-parameter model, cut trainable parameters by a factor of 10,000 and memory needs by a factor of three [10].

The domain data is the hard part. CISA notes that industrial datasets are hard to build well, with legacy systems and rare safety-related cases, and that operator expertise is needed to capture them [5]. It also suggests a data usage policy, since operators may not want vendors training on operational data that holds intellectual property [5]. In the EU, whoever fine-tunes may become a model provider; the Commission's indicative criterion is modification compute above one third of the original model's training compute [6].

## How should any of them be tested for engineering work?

NIST's Generative AI Profile (NIST AI 600-1, July 2024) warns that tests designed for humans, such as professional licensing exams, do not guarantee a model's validity in that field, and that benchmark results may not carry over to real-world conditions [4]. It adds that benchmark datasets can contain label errors that affect model selection [4]. FormationEval's author reports a residual bias toward longer correct answers in its own set [14].

**Analysis:**
The two O&G benchmarks above disagree by more than 20 percentage points at the top, on neighbouring subjects [14] [13]. A public score is a reason to shortlist a model, not to trust it. The test that counts is a set of the operator's own tasks, with known answers, run on the exact version that will be used and rerun when the version changes. The failure patterns to look for are set out in [why general AI tools get engineering answers wrong](/library/ai-errors-in-engineering/).

## A common mix-up

The usual mix-up is to read "open" as a mark of quality or safety. OSI's definition is about freedoms to use, study, modify and share [1]; accuracy and risk are measured separately, in the context of use [4]. Nor is a specialized model a checked one: fluent engineering terms can still carry confident errors, which NIST calls confabulation [4].

## What should an operator ask before choosing?

- **Where do the weights and the data sit?** Know where training and prompt data is stored, who can reach it, and whether access is remote, cloud or offshore [5].
- **What does the licence allow?** Read the use restrictions, naming rules and user thresholds, not only the price [8] [9].
- **Who changes the model, and when?** Fix the version under test and monitor behaviour after any update [12].
- **What is the supply chain?** Ask vendors where models are hosted and for a software bill of materials covering the AI [5].
- **How was it tested, on what?** Ask for evaluation on tasks like the operator's own, proportional to the risk of the use [4].
- **Who signs the result?** Keep a named engineer accountable for anything the model drafts (see [industrial AI agent](/library/industrial-ai-agent/)) [5].

## Sources

1. [The Open Source AI Definition, version 1.0, Open Source Initiative (OSI)](https://opensource.org/ai/open-source-ai-definition)
2. [Open Weights: not quite what you've been told, Open Source Initiative (OSI)](https://opensource.org/ai/open-weights)
3. [Dual-Use Foundation Models with Widely Available Model Weights (NTIA report), US National Telecommunications and Information Administration (NTIA)](https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf), 2024-07-30
4. [Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf), 2024-07-26
5. [Principles for the Secure Integration of Artificial Intelligence in Operational Technology (joint guidance, TLP:CLEAR), US Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre and partner agencies](https://www.cisa.gov/sites/default/files/2026-01/joint-guidance-principles-for-the-secure-integration-of-artificial-intelligence-in-operational-technology-508cV2.pdf), 2025-12-03
6. [General-Purpose AI Models in the AI Act: Questions and Answers (last update 9 September 2025), European Commission, Shaping Europe's digital future (europa.eu)](https://digital-strategy.ec.europa.eu/en/faqs/general-purpose-ai-models-ai-act-questions-answers), 2025-09-09
7. [Apache License, Version 2.0 (January 2004), The Apache Software Foundation](https://www.apache.org/licenses/LICENSE-2.0), 2004-01-01
8. [Llama 3.1 Community License Agreement, Meta Platforms (model owner's licence page)](https://www.llama.com/llama3_1/license/), 2024-07-23
9. [Llama 3.1 Acceptable Use Policy, Meta Platforms (model owner's policy page)](https://www.llama.com/llama3_1/use-policy/)
10. [LoRA: Low-Rank Adaptation of Large Language Models (E. J. Hu, Y. Shen, P. Wallis and others), arXiv:2106.09685, arXiv preprint (published at ICLR 2022)](https://arxiv.org/abs/2106.09685), 2021-06-17
11. [Don't Stop Pretraining: Adapt Language Models to Domains and Tasks (S. Gururangan, A. Marasović, S. Swayamdipta and others), arXiv:2004.10964, Proceedings of ACL 2020, preprint on arXiv](https://arxiv.org/abs/2004.10964), 2020-04-23
12. [How is ChatGPT's behavior changing over time? (L. Chen, M. Zaharia, J. Zou), arXiv:2307.09009, arXiv preprint](https://arxiv.org/abs/2307.09009), 2023-07-18
13. [PetroBench: A Benchmark for Large Language Models in Petroleum Engineering (X. Wang, T. Zhang and others), arXiv:2605.28032, arXiv preprint](https://arxiv.org/abs/2605.28032), 2026-05-27
14. [FormationEval, an open multiple-choice benchmark for petroleum geoscience (A. Ermilov), arXiv:2601.02158v2, arXiv preprint, CC BY 4.0](https://arxiv.org/abs/2601.02158), 2026-02-14

## From Winda

Winda's proprietary model is trained for the work from design to asset strategy, with engineers in the loop: engineers review the model's work, and their judgment trains it. It runs cloud, hybrid or air-gapped, as the operator chooses, and the data stays in the operator's account. [See Winda's specialized model](https://winda.ai/ai-platform/model/)

Published by Winda Beacon (https://beacon.winda.ai).
