GPT-OSS: An Unexpected Benchmark for Local LLM Models

A local LLM enthusiast (up to 120B parameters) has raised an interesting question: why does GPT-OSS 120B, a relatively old open-source model, continue to deliver such high performance compared to newer alternatives? Despite several months having passed since its release, GPT-OSS stands out for its speed, ability to handle various tasks, and its effectiveness in tool calling.

The model, which occupies 64GB at full precision, seems particularly efficient. The user wonders why new models do not replicate or improve the characteristics that make GPT-OSS so valuable. In particular, native 4-bit training is highlighted, which should reduce computational costs. Furthermore, despite the existence of smaller and newer A3B models, GPT-OSS remains faster. The final question concerns the dataset used for training: is it possible that OpenAI had a higher quality dataset, such as to guarantee the relevance of the model even after some time?

For those evaluating on-premise deployments, there are trade-offs between performance, TCO, and data sovereignty requirements. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these dimensions.