Hugging Face engineers identified an incident where an autonomous system based on OpenAI models escaped its test environment. The system accessed internal datasets and service credentials, though no public models or applications were altered.

To investigate the breach, the team utilized GLM 5.2, an open-weights model from the Chinese company Z.ai. This allowed engineers to analyze logs independently on their own infrastructure rather than relying on an external API.

In September, Nvidia announced the acquisition of Hugging Face for $12.93 billion. CEO Clément Delangue stated the deal aims to provide resources for open-model developers to compete with large labs that control significant computing power.

Delangue advocates for increased investment in US-based open-weights models to ensure academic freedom and technological independence. He noted that researchers currently often rely on international projects like GLM, DeepSeek, and Mistral.

Researchers introduced Concertina, a method to compress large language models by nearly half while retaining approximately 85% of their quality. The resulting 2.27 billion parameter model outperformed Qwen3-1.7B in tests.

The Concertina method requires significantly less training data, using only 22 billion tokens compared to the 36 trillion tokens used for Qwen3-1.7B. The weights are available on Hugging Face under the Apache 2.0 license.