Openai releases new open models

On Tuesday evening, Openai published two open-weight-reaing models with GPT-OSS-1220B and GPT-OS-20B. Openai boss Sam Altman had already promised the first openweight models since GPT-2 in March. In July, however, the publication was postponed from security concerns at an indefinite period. Unlike closed models such as GPT-4 or O4-Mini, open models also contain the trained parameters, which enables third parties to operate on their own infrastructure.

After Information from Openaai GPT-OS-12M20B in central benchmarks reaches the performance of Openais Reasoning model O4-Mini, but runs to just one 80 GB-GPU. GPT-OS-20B, however, provide strong starting skills on a 16 GB-EDGE device. Reasoning models are specifically trained to draw complex logical conclusions, solve multi-level problems and explicitly present their thinking processes.

New security protocol developed

Open models are also interesting for use in high -security environments because they can be operated locally. However, the safety requirements are also higher for the models, since this cannot be withdrawn after approval – unlike closed models. Openai boss Sam Altman apparently still had concerns in July and explained that additional security tests were necessary. In the course of the current publication, the company said 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 cyber security, should therefore aim to prevent the model for weapon construction or cyber attacks. External experts had checked the methodology and Openaai also makes evaluation code, prompts and evaluation guidelines public.

“Characterized by democratic values”

Together with the publication of the new models, development manuals and tools were also provided to adapt the models for your own needs and implement protective mechanisms. They are published under the Apache 2.0 license and can be integrated into Hugging Face, Vllm, Ollama and Llamaa.cpp, among others.

On the occasion of the publication, Openai boss Altman said that the world can “build on an open AI stack in the future-developed in the United States, shaped by democratic values, free of charge and for the broad benefit of everyone.” The statement can be understood as an allusion to Chinese Open Weight models that were very popular after its publication, but at the same time called critics on the scene. So the training material is coordinated to answer questions in the sense of the Chinese government. Historical events like on Tian’anmen-Platz hide the AI.