
News + Trends
ETH and EPFL present their own AI model
by Samuel Buchmann
GPT-OSS can be operated locally or in the cloud. The larger of the two variants is said to be roughly as powerful as OpenAI's proprietary o4-mini-LLM.
OpenAI has released GPT-OSS, an open-weight model for the first time in six years. It appears in two variants with 120 and 20 billion parameters and is freely available under the Apache 2.0 licence.
The larger version requires a GPU with 80 gigabytes of VRAM and, according to OpenAI, should perform similarly to the proprietary o4-mini model in benchmarks, which in turn is a small variant of GPT-4o - the model behind ChatGPT. The smaller version of the new GPT-OSS is designed for use on devices with 16 GB of VRAM and achieves results comparable to o3-mini. Both models are available on platforms such as Hugging Face, Databricks, Azure and AWS and can be customised for your own purposes.
According to OpenAI, GPT-OSS performs particularly well in the areas of reasoning, tool utilisation and chain-of-thought. The models were trained using current methods that are also used in OpenAI's advanced proprietary systems. A focus is also placed on security: GPT-OSS has been tested by external expert groups, among others, to minimise risks such as misuse in the area of cybersecurity or bioweapons.
GPT-OSS is thus positioning itself as an alternative to other open-weight models such as Meta's Llama 3 or DeepSeek, which can be operated both on consumer hardware and in the cloud. Open-weight models offer more transparency and control than proprietary models such as GPT-4o. However, training data and source code remain under lock and key.
A genuine open source counterpart to GPT-OSS comes from Switzerland: ETH Zurich and EPFL are working on a completely open language model that is due to be released in late summer 2025. It is based on up to 70 billion parameters. In contrast to GPT-OSS, source code and training data will be published alongside the weights. The Swiss model also emphasises compliance with data protection guidelines.
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.