Open vs Closed AI Models: How DeepSeek, Qwen and Other Developers Are Changing the Market and Lowering the Cost of AI Adoption

Just a few years ago, it seemed that frontier artificial intelligence would remain the privilege of a handful of companies with the largest compute budgets. Open models — above all DeepSeek and Qwen, along with Llama, Mistral, Gemma and other families — have changed that picture. They showed that comparable quality on many tasks is achievable at a much lower cost, and they set off price competition that benefits the companies using AI.

Key takeaways

• Open models have narrowed the gap with closed ones on many practical tasks;
• Their arrival has lowered the cost of using AI across the market, including closed APIs;
• "Open" does not mean "free": infrastructure and a team cost money;
• Closed models keep an edge on the hardest tasks and in service level;
• For most companies, the optimal strategy is a hybrid one.

What an open model is

In the industry, "open" usually refers to open-weight models: a company can download the model, run it on its own servers and fine-tune it for its own tasks. However, the training data and full training code are typically not disclosed, and the terms of use are set by a license. Some licenses allow virtually any commercial use, while others include restrictions, so a legal review is a mandatory step.

How DeepSeek and Qwen changed the market

DeepSeek drew widespread attention when it released models with strong reasoning abilities that were trained and served far more cheaply than the market expected. Engineering choices played a key role: a Mixture of Experts architecture, in which only part of the parameters is activated for each request, and an optimized training process. This forced the entire industry to rethink how much it actually costs to build a strong model.

Alibaba's Qwen family bet on the breadth of its lineup: from compact models that run on a laptop to large versions for complex tasks, plus specialized variants for code, math and image understanding. As a result, Qwen has become one of the most popular bases for fine-tuning in corporate and research projects.

The outcome is noticeable price pressure: closed-model providers are cutting prices and releasing lighter versions to stay competitive.

Where closed models are still ahead

• The hardest tasks. Multi-step reasoning, long-running autonomous agent work and complex programming are still more often handled better by flagship closed models.
• A ready-made service. APIs with availability guarantees, scaling, security tooling and support remove the need to build your own infrastructure.
• Speed of updates. A company receives model improvements without its own retraining and deployment effort.

The real total cost of ownership

The absence of a license fee does not make an open model free. The calculation has to include GPUs or cloud capacity, operations and machine learning engineers, monitoring, version upgrades and security. With small and uneven request volumes, a closed API often turns out to be cheaper. With large, stable workloads, strict data residency requirements or the need for deep customization, the advantage shifts to open models.

A hybrid strategy

More and more companies use several models at once. Simple, high-volume requests — classification, data extraction, knowledge-base answers — are handled by an inexpensive open model deployed in-house. Complex tasks are automatically routed to a flagship closed model. This kind of routing can cut costs substantially without sacrificing quality where it matters most.

Risks to consider

• license terms and restrictions on commercial use;
• regulatory requirements regarding model provenance and data processing;
• supply-chain security: models should be downloaded only from verified sources;
• the need for your own testing on corporate data, not just public leaderboards.

Bottom line

Competition between open and closed models is one of the main reasons AI adoption is getting cheaper. The choice is no longer about "which model is better." What matters more is understanding which combination of models delivers the required quality at an acceptable cost and level of control over data.

Prepared by the editorial team. This material is for informational purposes only.