Code sources and model weights for AI: Know their differences

Understanding model code sources in AI

Models are at the heart of every AI system. They learn patterns, enabling core AI capabilities such as interpretation, reasoning and decision-making. Models exist as code just like any other software, but they're uniquely challenging software entities. They demand extensive computing power and careful attention to vast quantities of high-quality data. In an ideal world, every AI system would be able to run its own models, but the nature of the AI project, the cost of the infrastructure, the availability of training data and the underlying skills of the AI project staff might not always support locally controlled models.

Consequently, the AI industry has generally split models into two broad categories. There are freely available open source models that businesses can download, modify, train, configure, operate using local or cloud infrastructures, and redistribute once modified. In comparison, third-party providers own, operate and maintain closed-source (proprietary) models, eliminating much of the hassle and overhead involved. The success or failure of an AI project can depend on the correct choice of model.

Open source AI models

Open source AI models resemble other open source software, with code open and available. They can be freely obtained -- often along with model parameters and weights. They can be modified by in-house software developers, enabling the business to tailor the model's behavior and performance. A business with nuanced model requirements almost always considers an open source AI model.

Open source models can also exhibit better technological independence than closed-source models because seeing, modifying, deploying, training, operating and managing the open source model eliminates the vendor lock-in and technological commitments that often accompany closed-source options. Further, having the source code, parameters and weights puts businesses in a stronger regulatory position, enabling closer scrutiny of security measures, bias monitoring, training data quality, transparency and explainability.

Benefits. When implemented correctly, open source AI models can run on local infrastructure, private clouds or public clouds. This enables businesses to scale their infrastructure investment in accordance with their AI computing needs. Running an open source model locally also ensures better data security because sensitive data never leaves the business. And, as with any open source product, a global community of enthusiastic developers can accelerate model improvements, bug fixes and new features.

Challenges. Open source models have their limitations. AI models are complex and impose significant infrastructure demands in networking, storage, computing and accelerators. Efficient AI model operation requires infrastructure with sufficient raw capacity for effective training and optimized inference. It's a costly investment. Software development demands and infrastructure requirements place an enormous burden on human skills, and businesses might struggle to hire or retain the skilled professionals needed. Finally, while open source AI models can provide better governance and security, they require tools, skills and active cross-functional collaboration to secure, operate and monitor AI models in-house.

Examples. Some examples of current open source AI models include DeepSeek-R1, Falcon-7B and Falcon-40B, OLMo and Qwen.

Closed-source AI models

Closed-source AI models are simply another type of proprietary software -- just like any other business application that a company might buy or subscribe to. The model's codebase, parameters and weights are the exclusive IP of their developer and are not made available to the public. In effect, a closed-source model is a black box that the provider owns and operates on its own infrastructure, allowing users to access the model using queries -- usually as an API or tokens -- and paying a per-token or per-API call fee for that access.

Closed-source AI models tend to be advanced and cutting-edge in their behavior, providing breadth and depth while delivering excellent overall performance. This makes closed-source AI models a good choice when general results and high performance are needed, without the requirement to run and manage the model locally.

However, closed-source models carry a high risk of vendor lock-in because users must accept the provider's exclusive decisions regarding model training, parameters, weights, versioning, performance and other elements. Further, the proprietary nature of closed-source models can obscure model transparency and make explainability difficult -- complicating AI governance and regulatory obligations for users.

Benefits. Closed-source AI models are particularly noted for their simplicity: they work well on demand with very little technical setup or specialized hardware. The models are also well supported by comprehensive documentation and readily available customer service options. Providers' infrastructure is usually strong and resilient, improving uptime and ensuring consistent performance under varying load conditions. Strong content moderation and guardrails can prevent undesirable or harmful outcomes.

Challenges. Closed-source AI models inevitably impose vendor lock-in. Users are committed to the provider's performance, responsiveness and future model roadmap. Disruptions to the provider's service can affect users, and migrating to another provider might be extremely difficult. There are serious data privacy issues when using third-party models; sensitive data might be sent outside the business, posing irreconcilable privacy risks, and the provider might also use that data to train its models further. The per-use cost of third-party AI models can increase rapidly with high utilization. Finally, the model's closed nature can make it impossible to determine why a mistake was made or why a bias occurred, which can carry serious regulatory consequences for the business.

Examples. Several examples of closed-source AI models include Anthropic Claude, Google Gemini and OpenAI GPT-4.

Understanding model weights in AI

Model weights are an array of parameters that affects how strongly certain inputs influence outputs. Weights effectively scale the values passed between the layers of a neural network. The larger the weight of a parameter, the greater the changes. In effect, weights control how an AI sees the relative importance of certain data inputs.

Weights are developed and strengthened as a by-product of training processes. The number of weights can vary dramatically across models: simple models might use few parameters, while advanced frontier large language models might involve billions or trillions of weights. It's possible to tune weights through a mathematical process called backpropagation, but retraining the model achieves a similar effect. Weights are typically not adjusted by direct manual setting, such as changing a number in a JSON file, because of unpredictable cause-and-effect on the model's predictive accuracy.

Weights are different from biases. Weights affect the strength and direction of the connection between nodes of a model's neural network. By comparison, biases provide an offset -- a plus/minus adjustment -- that helps the model fit data better. Biases and weights work together to provide the learnable parameters that define a model's behavior.

Open-weight AI models

Open-weight AI models have trained numerical parameters that are publicly available for download, modification and local use in AI systems. It's important to note that open-weight AI does not mean open source AI. Model developers might offer access to weights while keeping the model's source code and training data proprietary. This often provides developers a way to protect their IP investments while offering users greater flexibility to tune and optimize the model for unique AI needs.

Benefits. The most direct benefit of open-weight AI models is customization. Adopters can modify the model's behavior to fit specific business needs. Open-weight AI models can be well-suited to data privacy requirements, allowing organizations to fine-tune models using sensitive or personal data locally without sharing that information with third-party providers, where it might be disclosed or used to train other AI systems. This helps businesses maintain AI governance and regulatory compliance.

Challenges. Running models locally demands a significant investment in computing infrastructure using extensive memory and powerful processor acceleration techniques, making the model costly to run. Publicly available weights are helpful for transparency, but models that maintain proprietary source code and training data still face scrutiny regarding explainability. Finally, the model's provider cannot readily support it once weights are changed, making it difficult to find and fix bugs or prevent model misuse.

Examples. Common examples of open-weight AI models include DeepSeek, GLM series, Meta Llama, and Mistral AI.

Closed-weight AI models

Closed-weight AI models keep their weights, training data and source code completely hidden. This gives the provider full protection of its IP investments. The provider fully manages closed-weight AI models, and users access the model using APIs or tokens. In virtually all practical cases, closed-weight AI models share the same characteristics, benefits and challenges as closed-source AI models such as Anthropic Claude, Google Gemini and OpenAI GPT-4.

Emerging concepts in AI model availability

As AI models diversify and become more sophisticated, their availability strategies are also becoming more specialized. Some of the emerging trends in AI model availability include the following concepts:

• Source-available AI models. The idea of open source software typically involves licensing and restrictions that adhere to the Open Source Initiative (OSI). However, some providers don't adhere to OSI definitions, classifying the model's source code, weights and other details as source-available rather than open source. It's important to understand the subtle distinction because the provider's restrictions can affect how its model can be used or redistributed. Always read the fine print in any model licensing agreements.
• Ever-increasing model sophistication. Open source models have reached the trillion-parameter mark, offering incredible power and sophistication for agentic workflows.
• AI model decoupling. It's typically best practice to isolate the model from the associated AI application logic. The goal is to prevent vendor lock-in and enable rapid adoption of alternative models, but the workflow to ensure and support proper decoupling is often lacking. Emerging orchestration tools like n8n are deliberately designed to connect AI models to the overarching AI application, enabling clean isolation and smooth exchange of other models as desired.
• Tiered or restricted AI model access. This type of access means the provider controls who can use the AI model and how deeply users can interact with it in terms of speed, capabilities and permissions. These restrictions support additional control over security, cost and regulatory obligations. For example, Mythos 5.1 is reserved for defense firms.
• Increasing use of anti-distillation techniques. Knowledge distillation is a problem for model providers. It occurs when a provider invests in training a complex and expensive model (the teacher). Attackers query this model through its API and collect answers to fundamental reasoning questions. Those responses are then used to train smaller and simpler models (the students) that mimic the original model. Thus, student models steal IP from the teacher model. Anti-distillation techniques can alter its reasoning traces or shift its token patterns, leaving the responses useful to humans but useless to student models.

Stephen J. Bigelow, senior technology editor at TechTarget, has more than 30 years of technical writing experience in the PC and technology industry.