Amazon SageMaker AI introduces bidirectional streaming for real-time speech-to-text inference starting November 2025. Mistral AI's vLLM Realtime API allows for seamless bidirectional streaming between client and server for deploying compact real-time speech models, offering a fully managed, real-time speech-to-text service.
New MLLM-as-a-Judge evaluators in Strands Evals SDK enhance image-to-text tasks, predicting 80% enterprise software to be multimodal by 2030. Automated multimodal evaluation improves accuracy and efficiency in software development.
Kiro CLI introduces a custom Model Context Protocol server to enhance conversational memory with Amazon Bedrock AgentCore Memory, enabling AI agents to retain context for more intelligent interactions. The solution includes tools for searching, monitoring memory usage, and managing conversation history, improving productivity and personalization.
Amazon SageMaker Feature Store now supports Apache Iceberg format, streaming ingestion, and fine-grained access control through AWS Lake Formation, addressing operational challenges for ML models. New capabilities in SageMaker Python SDK v3. 8. 0 include Lake Formation integration, Iceberg table properties control, and modular Feature Store support, simplifying access control and reducing stora...
PTC reduces latency and token usage by allowing large language models to programmatically call multiple tools within a sandboxed environment, improving efficiency for multi-tool workflows. This innovative approach is particularly effective for data processing, numerical calculations, process orchestration, and privacy-sensitive tasks, offering a model-agnostic solution for improved performance ...
NVIDIA introduces NVFP4 for 4-bit floating point training, achieving 62.58% accuracy on Mamba-Transformer, surpassing FP8 baseline. NVFP4 optimizes dynamic range and precision, running GEMMs at 2-3x speedup over FP8 on Blackwell Tensor Cores.
Amazon Bedrock AgentCore Evaluations offers custom code-based evaluators for assessing agentic applications in specialized domains like financial services. These evaluators provide control over scoring logic, allowing for tailored assessments of agent quality and seamless integration into development workflows.
Language models face optimization challenges due to uneven token distributions. Adam's adaptive optimization helps rare tokens learn faster than with standard SGD.
AdaBoost regression uses decision trees trained on weighted data for better predictions. Results show overfitting with high accuracy on training data but lower accuracy on unseen test data.
MemPrivacy by MemTensor, HONOR Device, and Tongji University replaces private user data with structured tokens to protect privacy in cloud memory management without sacrificing utility or response quality. This local reversible pseudonymization framework ensures semantically intact interactions while safeguarding sensitive information from exposure in cloud systems.
Aderant streamlines support with Amazon Quick, achieving 90% faster search times and 75% documentation acceleration. AI-powered capabilities unify search across six systems, empowering engineers to deliver faster, more responsive support.
Learn how to fine-tune Amazon Nova for content moderation tasks using structured and free-form prompts. Benchmark Amazon Nova 2 Lite against foundation models on public datasets using the MLCommons AILuminate Assessment Standard.
FlashAttention addressed the expensive attention issue in large language models, but Lighthouse Attention by Nous Research achieves faster training with lower memory usage, challenging existing sparse methods. Lighthouse's innovative four-stage pipeline optimizes symmetric pooling and selection logic for improved performance in transformer models.
NVIDIA's SANA-WM tackles challenges in scaling world models for high-resolution video synthesis. It features a 2.6B-parameter Diffusion Transformer and supports single-GPU inference for fast deployment.
AI-driven search and chat help employees find answers in large repositories. Amazon Quick now offers document-level ACL support for fine-grained control over access to sensitive documents, ensuring compliance and data governance.