Bairo GonzalezLeandro · Martinez
Start your guided journey The story

In the world · Artificial intelligence · 5 August 2025

Open AI models: gpt-oss, DeepSeek and what changes

In 2025 DeepSeek opened R1 and OpenAI launched gpt-oss, its first open weights since 2019. Why open models matter for the small player.

bairogonzalez.com team, drawing on Bairo's story · Published

An open model, or open-weight model, is AI that can be downloaded, run on your own computer and adapted, instead of only being used over the internet. In 2025, two launches defined the topic: in January, China's DeepSeek published the reasoning model DeepSeek-R1 under the MIT license; on August 5, OpenAI launched gpt-oss, in two sizes, under the Apache 2.0 license, the company's first open weights since GPT-2, in 2019.

What happened

DeepSeek-R1 was published under a license that allows commercial use, modification and derivative works. Along with it came smaller, "distilled" versions, in sizes that fit on far more modest machines. The launch caught the world's attention because an open model came close to the best closed models of the time.

In August, OpenAI introduced gpt-oss, in two versions: gpt-oss-120b and gpt-oss-20b. According to MIT Technology Review:

  • the smaller version runs on a computer with 16 GB of memory;
  • performance is close to that of OpenAI's own small models on several tests;
  • the Apache 2.0 license allows commercial use and was seen by researchers as more open than that of other American models;
  • the company also presented the launch as a response to the advance of Chinese open models, such as DeepSeek, Qwen and Kimi.

The magazine highlighted a use that matters to many: hospitals, law firms and governments, which need to keep data in-house, can run and fine-tune the model locally.

Stanford's AI Index 2025 measured the effect: in one year, the performance gap between open and closed models fell from 8% to 1.7% on some benchmarks.

Why it matters

Open models change three things for those who are not giants:

  • Privacy. Sensitive data, such as medical records, case files or accounting, can be processed without leaving the company's machine.
  • Cost. For bounded tasks, a small open model can cost a fraction of what a large model costs over the internet.
  • Specialization. An open model can be fine-tuned for a niche, with the vocabulary and rules of that trade.

There are caveats. An open model still makes mistakes, security becomes the responsibility of whoever runs it, and each license has its own conditions. Opening the weights is not the same as opening the training data.

In Bairo's view

For Bairo Leandro Gonzalez Martinez, open models are one of the most concrete ways of giving AI back to people. In his reading, if AI is an extension of the human, it cannot depend only on those who rent it out. It needs to be able to live close to those who use it.

He says CLAIN was born from a calculation: AI consumes tokens, tokens cost money, and those who need it most are those who can least afford it. For him, small open models are part of the answer, because they make it possible to have inexpensive specialists, each trained for one concrete pain point. That is the design of CLAIN, the agent orchestrator he founded: many niche AIs, coordinated, with a person deciding at the end. The goal of 142 niche AIs in 142 days, announced on October 1, 2026, is a public goal.

Bairo also draws attention to the responsibility side. In his reading, opening the model broadens access, and precisely for that reason calls for more care with the truth of what comes out of it.

Where this meets the ecosystem

  • In health, Xperienc Global Labs works with data that requires confidentiality. The proposal is private records in the user's dashboard and a clinical assistant validated by a professional.
  • In law, SOSPROCESSOS, under development, deals with case information, a use in which running the model in a controlled environment makes a difference.
  • In file protection, qrqbits proposes layered post-quantum cryptography for files, images, audio and text.

Each is a distinct entity, with its own registration. The choice of which model to use in each case is technical and depends on evaluation.

Sources

Read also

More on this topic

In the world: Artificial intelligence