Hugging Face AI Updates: November 13, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. IBM Releases Granite 4.1 LLMON Preview Models, Trained on a Markup Language That Separates Instructions From Data
IBM’s Granite team published LLMON (LLM Object Notation) on the Hugging Face blog on November 12, together with two experimental models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, and a GitHub toolkit with lossless JSON converters. LLMON is an XML-like markup whose six structural delimiters are reserved special tokens, and each turn carries an execution block that binds an instruction to the data and requirements it applies to. On the OPI indirect prompt injection benchmark, IBM reports the previews avoid 47% more attacks than the best frontier model tested and 95% more than the best open source model, with general benchmark results on par with Granite 4.1 Instruct. IBM describes the separation as learned behavior and not an enforced guarantee, and says the preview is not intended for production. Source
2. Hugging Face Publishes a LoRA Fine-Tuning Guide for TRELLIS.2 With Docker Images for Hugging Face Jobs
Hugging Face published a guide on November 12 with scripts and two container images (preprocessing and training) for fine-tuning the TRELLIS.2 image-to-3D model, runnable on Hugging Face Jobs or locally with Docker. The example trains about 9.3 million LoRA parameters with PEFT on 188 footwear objects from Google Scanned Objects: rank 8 adapters on the sparse-structure flow DiT and rank 16 on the shape flow DiT, with the texture flow DiT frozen because the dataset lacks separate material maps. The two training runs took 7.2 and 3.6 hours, each on a single A100 80 GB, and on nine held-out objects the mean F-score at 0.02 rose from 0.276 to 0.803 while Chamfer distance fell from 0.174 to 0.033. Source