1. Closed vs Open Weight LLMs

Understanding the spectrum between managed frontier APIs and self-hosted models—and choosing the right foundation for your AI agent.

When building an AI agent, one of the most important architectural decisions is choosing the model—or combination of models—that will power it. This is often described as a choice between proprietary, closed-weight models and open-weight models.

In practice, the distinction is no longer a simple two-way divide. Models now sit on a spectrum that includes fully managed proprietary APIs, downloadable models with restrictive community licences, permissively licensed open-weight models, and more fully open AI systems that publish additional training code and data information.

A model’s weights are the numerical parameters learned during training. They encode the patterns the model uses to interpret inputs and generate outputs. With a closed-weight model, those parameters remain under the provider’s control. With an open-weight model, the model files can be downloaded and deployed by others, subject to the model’s licence.

Open weights should not automatically be confused with open source. A model may provide downloadable weights while withholding its training data, data-processing pipeline, training code, intermediate checkpoints, or other information needed to reproduce it. Always examine the model’s licence and documentation rather than relying solely on labels such as “open,” “open source,” or “open AI.”

Managed and Closed-Weight Models

Closed-weight models are operated by the organizations that developed them. Developers normally access them through an API or managed cloud platform rather than downloading the model itself.

Your application sends a request to the provider, the provider runs the model on its infrastructure, and the resulting response is returned to your application. Some providers also supply managed tools for web search, file retrieval, code execution, computer use, voice, image processing, structured outputs, and agent orchestration.

Leading Managed Model Families

Model catalogues change frequently. Always check the provider’s current documentation for availability, context limits, tool support, regional access, pricing, and deprecation dates before selecting a production model.

Open-Weight Models

Open-weight models make their trained parameters available for download. Depending on the licence and model size, they can be run on a developer workstation, an on-premises GPU server, a private cloud, a specialist inference provider, or a major public-cloud platform.

Open weights provide greater deployment flexibility, but the degree of openness varies substantially. Some models use permissive licences such as Apache 2.0 or MIT. Others use custom community licences with usage, redistribution, attribution, or scale-related restrictions.

Leading Open-Weight Model Families

The Gap Is No Longer One-Dimensional

It is increasingly inaccurate to say that all closed models are more capable or that all open-weight models are cheaper. The answer depends on the task, deployment scale, hardware, latency target, required modalities, context length, safety requirements, and the quality of the surrounding agent system.

A relatively small open-weight model can outperform a much larger frontier model on a narrow, well-defined workflow after careful fine-tuning or retrieval augmentation. Conversely, a managed frontier model may be far more economical when usage is intermittent because you pay only for requests and avoid maintaining idle GPU capacity.

For agents, model quality is only one part of reliability. Tool schemas, retrieval quality, permissions, state management, validation, retries, observability, sandboxing, and human approval gates can matter as much as the underlying model.

Making the Choice for AI Agents

There is no longer a universal rule that every team should prototype with a closed model and later replace it with an open-weight model. A better approach is to evaluate both against the actual workflow.

Choose a Managed Model When:

• You need access to the strongest available reasoning, coding, multimodal, or computer-use capabilities.
• You want to reach production quickly without operating GPU infrastructure.
• Your workload is unpredictable, intermittent, or too small to keep dedicated accelerators well utilized.
• You need managed tools, enterprise support, regional hosting, or contractual data controls.
• Your team does not have specialist model-serving and machine-learning infrastructure expertise.

Choose an Open-Weight Model When:

• You require on-premises, offline, edge, sovereign, or tightly controlled private-cloud deployment.
• Your data-governance policy does not permit external model processing.
• You need to customize the model beyond the options exposed by a managed API.
• Your workload is stable and large enough to justify dedicated inference infrastructure.
• You need long-term control over model versions and want to reduce exposure to API deprecations or provider policy changes.

Consider a Hybrid or Multi-Model Architecture

Many production agents now use multiple models rather than a single universal “brain.” For example, a system might use:

• A small local model for classification, routing, extraction, or sensitive-data filtering.
• A fast managed model for routine tool calls and conversational responses.
• A frontier reasoning model only for difficult planning or exception handling.
• A specialist coding, vision, speech, embedding, or reranking model for a specific stage of the workflow.

This routing approach can improve cost, latency, privacy, and resilience. It also makes it easier to replace an individual model as the market changes.

A Practical Evaluation Process

Before selecting a model, build a representative evaluation set from the tasks your agent must actually perform. Test candidate models on:

• Task completion and factual accuracy.
• Tool selection and argument generation.
• Structured-output reliability.
• Long-context performance and retrieval quality.
• Latency, throughput, and total cost per successful task.
• Security, prompt-injection resistance, and permission handling.
• Behaviour across retries, failures, and ambiguous instructions.
• Licence, privacy, retention, residency, and compliance requirements.

Measure the cost of a successful end-to-end task, not merely the advertised price per token. A cheaper model that needs repeated retries, longer prompts, additional validation, or frequent human intervention may be more expensive overall.

Further Reading & Resources

• Hugging Face Models — A major catalogue for discovering, comparing, and downloading open-weight models.
• LMArena — A crowdsourced comparison platform based partly on human preference evaluations.
• Artificial Analysis — Independent comparisons covering model quality, speed, context, and API pricing.
• Open Source AI Definition — The Open Source Initiative’s framework for distinguishing open-source AI systems from models that merely publish their weights.
• llama.cpp — A widely used runtime for running quantized models on local and edge hardware.
• vLLM — A popular high-throughput inference and serving engine for open-weight models.

Last reviewed: August 2026. Model names, licences, prices, context limits, and availability change quickly. Verify the current provider documentation and licence terms before making an architectural or procurement decision.