Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 5, 2026, 9:00 AM
Hugging Face blog details the training engineering implementation of the multilingual large language model BLOOM (176B parameters): using 384 NVIDIA A100 80GB GPUs on France's Jean Zay supercomputer, it took approximately 3.5 months to complete.
The author built a text-driven playlist generator based on Sentence Transformers and Gradio, using the ms-marco-MiniLM-L-6-v3 model to segment lyrics by paragraph and encode them into embedding vectors.
Hugging Face has released BLOOM, the world's largest open-source multilingual large language model, containing 176 billion parameters and supporting 46 natural languages and 13 programming languages; the model, intermediate training checkpoints, optimizer states, and inference API are all open. It was trained for 117 days on France's Jean Zay supercomputer, led by the BigScience project, and uses the Responsible AI License.
Hugging Face has launched Evaluation on the Hub, a zero-code model evaluation tool based on AutoTrain, which supports evaluating any combination of models and datasets on the Hub.
The official Hugging Face blog introduces how to seamlessly integrate DeepSpeed ZeRO (Stage-2 and Stage-3 CPU offload) via the Accelerate library for large model training.
OpenAI trained a neural network using Video PreTraining (VPT) on unlabeled human gameplay videos of Minecraft. With only a small amount of labeled contractor data for fine-tuning, the model can complete diamond tool crafting tasks that require skilled players more than 20 minutes (24,000 steps). The model directly uses keyboard and mouse inputs, offering versatility and representing a step toward general-purpose computer-use agents.
The official Hugging Face blog published an introductory tutorial on embeddings, using the U.S. Social Security Administration's Medicare FAQ as an example, demonstrating in three steps: 1) Calling the Hugging Face Inference API (using the sentence-transformers/all-MiniLM-L6-v2 model) to encode 13 questions into 384-dimensional embedding vectors.
Hugging Face and Intel have jointly launched the open-source library Optimum Intel, which implements optimizations such as quantization and pruning for Transformer models based on Intel Neural Compressor. The article provides a complete code example for post-training dynamic quantization of DistilBERT, showing that accuracy decreased from 0.574 to 0.546, inference time was reduced with a 1.34x speedup, and training is supported on Habana Gaudi accelerators.
Hugging Face Hub has introduced Pull Requests and Discussions features, now available in the Community tab across all repository types (Models, Datasets, Spaces). Users can directly create discussions and submit PRs without forking, achieved through custom refs stored in the source repository. This design is optimized specifically for ML repositories, distinguishing it from general-purpose Git platforms like GitHub.
The Hugging Face blog introduces the TAPEX method proposed by Microsoft, which constructs a synthetic pre-training corpus by systematically sampling executable SQL queries and their execution results on tables, enabling language models (such as BART) to learn neural SQL execution capabilities.
Hugging Face officially released Gradio 3.0, featuring a comprehensive frontend refactor (based on Svelte), new components such as Gallery, the TabbedInterface class, and a brand-new low-level Blocks API that supports building customized ML demo applications with multi-step workflows, dynamic layouts, and conditional rendering in Python.
Hugging Face Optimum 1.2 officially released its inference acceleration capabilities, supporting the acceleration of Transformers pipelines via ONNX Runtime.
The official Hugging Face blog announced that the Hub platform has officially integrated fastai, supporting the upload and loading of fastai Learners via the push_to_hub_fastai and from_pretrained_fastai functions.
Hugging Face has launched carbon emission-related features on the Hub: it supports filtering models by training CO2 emissions (in grams) via the emissions_threshold parameter, for example, a maximum of 100 grams or a minimum of 500 grams; simultaneously, codecarbon is integrated into the transformers library, automatically generating an emissions.csv file to record carbon emission data when training with Trainer, which can also be written to the model card.
Hugging Face announced the integration of the Decision Transformer (an offline RL method that models reinforcement learning as a sequence generation task) into the 🤗 transformers library.
The Hugging Face official blog published a tutorial demonstrating how to fine-tune the SegFormer-B0 model using a custom sidewalk image dataset (segments/sidewalk-semantic).
The official Hugging Face blog introduces how to use the datasets library to load the British Library's ancient book illustration dataset (embellishments_sample) and generate image embeddings using the CLIP-ViT-B-32 model from sentence_transformers.
Hugging Face's official blog published a tutorial demonstrating how to accelerate BERT text classification inference using Hugging Face Transformers, Amazon SageMaker, and AWS Inferentia.
OpenAI releases new versions of GPT-3 and Codex, supporting editing or content insertion into existing text, no longer limited to text continuation.
Hugging Face has added the constrained beam search feature to the Transformers library, supporting forced generation of specified words through force_words_ids or custom Constraint objects.
OpenAI built a neural theorem prover for Lean that successfully solved various high-difficulty high school math olympiad problems, including AMC12, AIME, and two problems adapted from the IMO.
The Hugging Face blog introduces how to leverage the CTC architecture characteristics of Wav2Vec2 to perform long audio or real-time speech recognition on limited hardware by using chunking with stride (chunk_length_s and stride_length_s), thereby avoiding GPU memory overflow; the article provides directly runnable pipeline code examples and explains that this method is also applicable to enhanced models integrated with language models.
OpenAI trained language models that are better at following user intent, more truthful, and less toxic than GPT-3, known as InstructGPT; these models were trained using alignment techniques such as reinforcement learning from human feedback and have been deployed as the default language model on the OpenAI API.
OpenAI has launched a new embeddings API endpoint that supports natural language and code tasks, including semantic search, clustering, topic modeling, and classification.