Today's highlights include the introduction of the Ettin Reranker family, which enhances retrieval-augmented generation tasks with state-of-the-art performance on benchmark datasets like MS MARCO and TREC (). Additionally, Granite 4.1 has been released, featuring large language models optimized for efficiency with up to 70 billion parameters, improving deployment in production environments (). AWS has also introduced new tools for foundation model training and inference, streamlining the deployment of large language models (). These advancements are crucial for practitioners looking to enhance the performance and efficiency of AI applications.
The Ettin Reranker family has been introduced, featuring models designed to enhance retrieval-augmented generation tasks. These models leverage a transformer-based architecture with improvements in fine-tuning techniques, achieving state-of-the-art performance on benchmark datasets such as MS MARCO and TREC. This release provides practitioners with robust tools for improving search relevance and information retrieval efficiency in AI applications.
Hugging Face Blog — 57 d ago · found 33 d agoModels2 · 1 cmts
2.
Granite 4.1 LLMs: How They’re Built
Granite 4.1 introduces a new family of large language models (LLMs) optimized for efficiency and performance across various tasks. Key enhancements include a model size of up to 70 billion parameters, improved fine-tuning techniques, and a redesigned transformer architecture that reduces latency while maintaining high accuracy on benchmarks such as GLUE and SuperGLUE. This release is significant for practitioners as it provides a more resource-efficient option for deploying LLMs in production environments, enabling broader accessibility and scalability in AI applications.
Hugging Face Blog — 76 d ago · found 33 d agoModels1 · 0 cmts
3.
Building Blocks for Foundation Model Training and Inference on AWS
AWS introduced a suite of tools and services designed to streamline the training and inference of foundation models, including optimized instances for large-scale model training, pre-built model architectures, and integrated frameworks for distributed training. Key features include support for popular frameworks like TensorFlow and PyTorch, as well as enhancements in Amazon SageMaker for deploying large language models with reduced latency. This development is significant for practitioners as it simplifies the infrastructure setup and accelerates the deployment of state-of-the-art AI models in production environments.
Hugging Face Blog — 64 d ago · found 33 d agoTraining
the full briefing
Models & Releases
The Ettin Reranker family has been introduced, featuring models designed to enhance retrieval-augmented generation tasks. These models leverage a transformer-based architecture with improvements in fine-tuning techniques, achieving state-of-the-art performance on benchmark datasets such as MS MARCO and TREC (). Granite 4.1 introduces a new family of large language models (LLMs) optimized for efficiency and performance across various tasks, with a model size of up to 70 billion parameters and improved fine-tuning techniques (). AWS has also released a suite of tools designed to streamline the training and inference of foundation models, including optimized instances for large-scale model training and integrated frameworks for distributed training ().
Research
The Open ASR Leaderboard has been updated to include the Benchmaxxer Repellant model, which demonstrates significant improvements in automatic speech recognition (ASR) tasks, achieving a Word Error Rate (WER) reduction of 15% on the Common Voice dataset (). The vLLM framework has released version 1.0, focusing on improving the correctness of reinforcement learning (RL) algorithms before implementing corrective measures (). Additionally, Granite Embedding Multilingual R2 has been released under the Apache 2.0 license, offering multilingual embeddings with a context size of 32,000 tokens (Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality).