OpenAI publishes new open models
With two open-weight reasoning models, OpenAI apparently wants to stand up to the competition from China in particular. One challenge was security.
(Image: kovop/Shutterstock.com)
OpenAI published two open-weight reasoning models, gpt-oss-120B and gpt-oss-20B, on Tuesday evening. Back in March, OpenAI CEO Sam Altman announced the first open-weight models since GPT-2. In July, however, the release was postponed indefinitely due to security concerns. Unlike closed models such as GPT-4 or o4-mini, open models also contain the trained parameters, which allows third parties to run them on their own infrastructure.
According to OpenAI, gpt-oss-120B achieves the performance of OpenAI's reasoning model o4-mini in central benchmarks, but runs on only a single 80 GB GPU. gpt-oss-20B, however, delivers strong reasoning capabilities on a 16 GB edge device. Reasoning models are specially trained to draw complex logical conclusions, solve multi-level problems, and explicitly represent their thought processes.
New security protocol developed
Open models are also interesting for use in high-security environments, as they can be operated locally. However, the safety requirements for the models themselves are also higher, as they cannot be taken back after release – unlike closed models. OpenAI boss Sam Altman apparently still had reservations in July and explained that additional security tests were necessary. In the course of the current release, the company explained that the models had been developed with a new security standard.
This "worst-case fine-tuning" protocol simulates malicious use in the areas of biology and cybersecurity, and should therefore aim to prevent the model from being misused for weapons manufacturing or cyberattacks. External experts have reviewed the methodology and OpenAI is also making evaluation code, prompts and evaluation guidelines public.
Videos by heise
"Shaped by democratic values"
Along with the publication of the new models, development manuals and tools have also been made available to adapt the models to your own needs and implement protection mechanisms. They are published under the Apache 2.0 license and can be integrated into Hugging Face, vLLM, Ollama and llamaa.cpp, among others.
OpenAI CEO Altman explained on the occasion of the release that the world "can build on an open AI stack in the future – developed in the USA, shaped by democratic values, accessible free of charge and for the broad benefit of all." The statement can be understood as an allusion to Chinese open-weight models, which enjoyed great popularity after their publication, but at the same time drew criticism. The training material is designed to answer questions in the interests of the Chinese government. The AI does not mention historical events such as Tiananmen Square.
(mki)
