IBM has released two experimental AI models that read instructions and reference material separately. The goal is to keep hidden malicious commands in documents or tool outputs from changing what the AI is supposed to do. The models were published on the Hugging Face technical blog.
Jensen Huang's resignation as Nvidia CEO has raised questions about whether the company will keep its open AI strategy. The planned Hugging Face acquisition has also not yet closed.
After Nvidia CEO Jensen Huang stepped down, Hugging Face CEO Clément Delangue posted that he will keep fighting for open-source AI. Nvidia plans to acquire Hugging Face, and the U.S. Justice Department is looking into how the deal would affect competition.
Hugging Face released a guide for training the 3D image generation model TRELLIS.2 on your own data. It includes fine-tuning scripts and ways to run them. The goal is to fix the mistakes a general model makes on specific objects.
Chinese AI company Moonshot AI has published the weights of its large AI model, Kimi K3.1 Pro, on Hugging Face, a site where AI models are shared. With the weights, companies can run and modify the model on their own servers.
Kakao has released its in-house decision-support model, Kanana-2-v-decide, on Hugging Face. The company says each request takes an average of 0.04 seconds or less to process. The model produces results in fixed formats such as probabilities and scores rather than sentences, so services can use them right away.
Kakao has released 'Kanana-2-v-Decide', a multimodal model that helps AI agents make decisions, on the open-source platform Hugging Face. The model looks at both text and images and reports results in a set format.
Kakao has released its own multimodal model, Kanana-2-v-decide, as open source on Hugging Face to help AI make decisions. Multimodal refers to AI that understands text and images together. Instead of full sentences, this model produces decisions in a fixed format.
Egyptian-accented and Nigerian-accented English speech data have been added to the Open ASR Leaderboard, Hugging Face's ranking of speech recognition performance. The goal is to fairly evaluate speech recognition (technology that turns spoken words into text) across a wider range of accents. This was a collaboration with Liva AI.
Moonshot AI has released the weights (the learned numerical values that make up a model) of its AI model Kimi K3.1 Pro on Hugging Face. This means developers can run and modify the model on their own servers. The release also revealed that its internal structure incorporates DeepSeek's sparse attention technology.
The Korea Electronics Technology Institute (KETI) has released four datasets for building and testing Korean AI. They cover conversation, shopping consultations, detection of false answers, and safety evaluation. Anyone can download them from AI Hub and Hugging Face.
The Hugging Face Transformers team found it hard to run tests on large models after every change, so it proposed a way to check quickly using small models. The effort was prompted by a case in late July 2026, when a bug in the UMT5 model that let it peek at future tokens stayed in the main branch for about two months.