His model is not only powerful and free, but it has cost much less

Openai yesterday launched its first models with open weights since 2019: GPT-Oss-120B and GPT-Oss-20B. Two reasoning systems that mark a radical change in the company’s strategy after Deepseek R1’s success At the beginning of the year.

Why is it important. Openai has gone from defending the safety of closed models to recognize that “it was on the wrong side of history” with open models. The earthquake that caused Deepseek In January he forced them to rethink their entire strategy.

The context. Until now, Openai had kept all its models closed since GPT-2justifying it for security reasons. But rivals as a goal with Calls And especially Chinese companies with Depseek, Qwen and Kimi They have gained ground in the open ecosystem.

The turning point came when Depseek R1 showed that it was possible to create advanced reasoning models with a fraction of the cost of Americans. Sam Altman then admitted that Openai needed “finding a different open source strategy.” And here it is.

In figures:

• The major model, GPT-Oss-120bworks at a single 80 GB GPU with just 5,100 million active parameters of its 117,000 million total.
• His little brother, GPT-Oss-20bit can be executed in a laptop with 16 GB of memory (although the speed seems to be very slow with such a configuration, in the absence that we can try it more thoroughly).

The current situation. The Benchmarks independent They place GPT-Oss-120b with a score of 58 in the intelligence indexsurpassing O3-mini but being below O3 and O4-mini.

In competitive programming tasks it reaches 2622 points in Codeforces, while Depseek R1 achieves less score with a much heavier model.

In detail. Architecture uses a system of Mix of experts (Moe, Mixture of experts) that activates only a fraction of the total parameters for each consultation. This allows the 120,000 million model of parameters to activate only 4.4% in each consultation, while the 20,000 million activates 17.2%.

Both models maintain chain reasoning ability without direct supervision, following the principle that visible reasoning is necessary to detect undue behaviors.

Yes, but. OpenAI has not made public training data, maintaining part of the secret. Nor does it include multimodal capabilities, limited only to text.

• Hallucinations are significantly higher than in their own models: 49% and 53% respectively compared to 16% of O1.

Between the lines. This launch responds to both Chinese competitive pressure and the new Trump administration, which has urged US companies to release more technology to “promote the global adoption of aligned with American values.”

The company has worked with partners such as Orange, Snowflake and AI Sweden for real use cases, seeking to validate whether the advantages justify continuing to invest in open models.

Deepen. The models are available under Apache 2.0 license on platforms such as Hugging Facewith native support in tools such as Ollama, LM Studio and Services Cloud of the main suppliers. Openai has organized a network Teaming Challenge with $ 500,000 in prizes to identify security risks.

This is OpenAi’s response to his Deepseek moment: recognizing that open innovation is not only inevitable, but necessary to stay competitive in a panorama where cost and accessibility matter as much as performance.

In Xataka | The GPT-5 paradox: OpenAi needs the greatest jump in the history of AI just when it seems most impossible

Outstanding image | Solen FeyissabDima Solomin