Amazon SageMaker AI has implemented a new integration with MLflow, aimed at improving the management of machine learning models. This integration leads to data-driven optimization and benchmarking of AI models.
The key feature of this integration is the real-time streaming of benchmark results, which is expected to reduce data silos and streamline workflows.
SageMaker AI has launched a new user interface for generating inference recommendations, making it easier to deploy generative AI models. This UI simplifies model deployment by reducing setup time from hours to minutes.
The integration additionally enhances the ability to monitor discriminative machine learning models, aiding in tracking data drift and maintaining accuracy. This is crucial for applications susceptible to changing external factors.
Active monitoring helps detect issues early, providing greater reliability and accuracy in classifications and regression use cases.
This comprehensive update to SageMaker AI helps reduce time and expertise required for deploying and maintaining AI models. By simplifying complex processes, Amazon aims to make AI technologies more accessible and maintainable across sectors.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's briefNew every morning, and the back catalogue is archived by date.
Amazon SageMaker HyperPod Slurm now includes managed cluster extensions for observability and user management. This update allows administrators to declaratively manage software and settings on running nodes, reducing manual configuration drift in shared Slurm clusters for foundation model development.
Amazon SageMaker HyperPod launched the Global Inference Gateway, a SageMaker-managed service that routes inference requests across multiple HyperPod clusters, including those in different AWS Regions. This gateway automatically directs requests to available GPU capacity, addressing issues of uneven load and manual failover across distributed model serving infrastructure. It builds upon the existing Inference Gateway, which handles routing within a single cluster.
Amazon SageMaker Inference received several updates in 2026, including new routing strategies and broader framework support for HyperPod Inference, and deeper AWS AI integration for managed endpoints. Inference recommendations are now accessible from SageMaker AI Studio, the SageMaker Python SDK, and coding agents, allowing endpoints to route requests across Regions and accounts. These changes address challenges in model configuration, GPU utilization, and multi-region deployment for large language models.
Amazon SageMaker AI now automatically monitors generative AI inference endpoints for performance and cost inefficiencies. The new feature detects latency drift, unused GPU capacity, and missed optimization opportunities, notifying users via AWS Health to suggest improvements.
Amazon SageMaker AI now offers cross-Region inference routing, a new feature that automatically directs inference requests to available endpoints across different AWS Regions and accounts. This allows applications to maintain model availability during regional capacity shortfalls without requiring custom routing logic.
PolyAI's Dialog-RSN-1, a reasoning model for real-time complex dialog, is now available on AWS Marketplace and Amazon SageMaker JumpStart. This allows AWS customers to deploy the model within their own AWS environment for applications requiring live call understanding and real-time responses.
Amazon SageMaker AI Serverless Model Customization now supports Full Fine-Tuning (FFT) for large language models. This allows for superior model quality by updating all model parameters, overcoming limitations of parameter-efficient methods like LoRA, without requiring users to manage GPU infrastructure.
Amazon SageMaker JumpStart launched a new Model Leaderboard within SageMaker Studio, allowing users to compare over 70 open-weight foundation models. This tool provides side-by-side comparisons of model quality, speed, and cost, using standardized benchmarks and performance metrics measured by AWS. The leaderboard aims to simplify model selection by offering consistent data for various large language models from providers like Meta, Google, and IBM.
Salesforce's Agentforce Foundation team adopted cross-Region inference routing in Amazon SageMaker AI to manage model serving across 30+ global regions. This solution addresses capacity fragmentation and improves reliability by dynamically routing inference requests to available endpoints, while maintaining data residency commitments.
Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo, an open-source distributed inference framework, to improve the efficiency of serving large language models (LLMs). This integration addresses challenges like large model sizes exceeding single-instance GPU memory and slow token generation from long prompts by enabling multi-node deployment and disaggregated prefill and decode. The new capability allows for better resource utilization and scaling for LLM inference workloads on SageMaker HyperPod.
Amazon SageMaker AI real-time endpoints now support KV cache-aware routing, a new strategy that directs requests to instances with existing cached matches for prompts. This feature optimizes performance for large language models (LLMs) handling multi-turn conversations or long shared prompts by reducing redundant computation when GPU memory is full or conversations evolve.
Amazon SageMaker now includes new AI data preparation capabilities to help users build high-quality datasets for model development. These features allow for generating synthetic examples, assessing and improving existing datasets, and integrating expert feedback, addressing common bottlenecks in AI project timelines.
Aionic Labs' OpenTSLM, an open-source family of time-series language models (TSLMs), is now available on Amazon SageMaker JumpStart. These models interpret operational measurements and contextual data to help organizations identify emerging problems and understand system changes, offering a specialized approach to time-series analysis compared to general-purpose language models.
GS Caltex, an energy and chemical company, has integrated Amazon SageMaker Canvas with Amazon SageMaker Pipelines to streamline its MLOps workflow. This integration allows process engineers to build models in Canvas and ML engineers to deploy them into production using Pipelines, reducing deployment time from a week to a few hours. The company also reports lower inference costs by using batch inference every 30 minutes instead of continuous endpoints.
Amazon SageMaker HyperPod now includes a Data Capture capability, which records live inference inputs and outputs to Amazon S3. This feature allows teams to detect model drift, identify failure patterns, and use real-world interactions to improve deployed foundation models.
AWS has launched usage-based metering for Amazon SageMaker AI bidirectional streaming inference, allowing model providers to charge for actual consumption during live streams on AWS Marketplace. This change addresses billing challenges for real-time speech AI, where session duration and usage vary, enabling providers to align billing with customer consumption.
Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU (MIG) for model endpoints, allowing users to partition physical GPUs into independent units. This enables deploying more models per GPU and selecting fractional GPU resources, addressing previous limitations where models required a full accelerator regardless of actual needs.
This post explains how to use sequence-level knowledge distillation (SeqKD) to fine-tune AI shopping agents on Amazon SageMaker. This method transfers the behavior of a larger, more expensive "teacher" model to a smaller, more cost-effective "student" model, maintaining quality while reducing operational costs for agentic applications.
AWS Batch now includes quota management for SageMaker Training jobs, allowing administrators to divide compute capacity into quota shares for different teams or projects. This feature enables resource-sharing strategies like lending and borrowing idle GPU capacity to improve utilization and manage multi-team training environments more efficiently.
Amazon SageMaker AI Training and processing jobs now support Seekable OCI (SOCI) indexing, allowing container images to lazy-load. This change reduces container image download time, enabling jobs to start in seconds instead of minutes and decreasing GPU idle time by up to 90%.
Exowatt developed a repeatable AI pipeline using Amazon SageMaker serverless model customization to fine-tune models for its solar energy systems. This pipeline produces task-specific models for edge and cloud deployments, achieving higher accuracy than frontier models at significantly lower costs. The new process reduces the time to develop and deploy these models from months to days, enabling scalable AI operations for their growing fleet of solar units.
Amazon SageMaker Python SDK v3 now supports fleet-wide computer vision model lifecycle management at the edge, replacing SageMaker Edge Manager with SageMaker Neo for optimization. This update addresses the challenge of managing thousands of fine-tuned models for individual cameras in large-scale industrial deployments. The solution uses a dual-branch SageMaker Pipeline for training and evaluation, including device-optimized artifact compilation and AWS Step Functions orchestration.
Amazon SageMaker AI supports Multi-Turn Reinforcement Learning (MTRL) to train models for sequential, agentic tasks where feedback is only available at the end of a sequence of actions. This method addresses limitations of single-turn reinforcement learning and supervised fine-tuning for complex, multi-step processes. The approach allows training smaller, specialized models for business-specific workflows.
AWS introduced a reusable CDK construct with CloudFormation custom resources to automate SageMaker AI endpoint auto scaling during updates. This construct addresses deployment failures that occur when updating SageMaker endpoints with registered auto scaling, allowing configuration and scaling parameter changes in a single deployment. The solution prevents manual intervention and ensures auto scaling remains active after updates.
Amazon SageMaker AI launched serverless model customization with the general availability of training sessions. This update allows users to customize open-weight models using post-training techniques without managing GPU infrastructure, providing control over the training loop while SageMaker handles resource allocation and fault recovery.
Amazon SageMaker AI launched new training session APIs, allowing developers to write custom training loops for large language models (LLMs) without managing underlying GPU infrastructure. This update provides more control over fine-tuning algorithms, such as custom loss functions or reinforcement learning reward functions, while SageMaker handles operational aspects like distributed execution and fault recovery.
Amazon SageMaker AI now supports streaming low-latency speech from AI agents using the Kokoro-82M text-to-speech model. This integration allows audio playback to begin as soon as an agent generates each sentence, reducing wait times for users interacting with AI voice agents.
Amazon SageMaker AI now offers a method to optimize auto scaling targets for Text-to-Speech (TTS) endpoints, replacing guesswork with empirical measurements. This change addresses issues where traditional auto scaling metrics for LLMs or short requests are unsuitable for TTS, preventing over-provisioning or service degradation during traffic spikes.
AWS now allows new Amazon SageMaker AI customers to access a pre-configured SageMaker Studio environment quickly after signing up, using existing identities like Google or GitHub. This change removes the need for a full AWS account setup and credit card entry for most users, reducing friction for developers starting with SageMaker.
Amazon SageMaker JumpStart expanded its catalog by over 100 new models from 30 different providers between January and September 2026. This expansion provides users with more options for foundation models, addressing increased demand and the rapid release of open-source models.
Meta's Muse Spark Open, a coding-optimized open-weight model, is now available in Amazon SageMaker JumpStart. This integration allows developers to deploy the model for agentic coding workflows with a 1 million-token context window, supporting tasks like multi-file refactors and debugging.
Amazon SageMaker JumpStart now offers simplified model discovery and deployment through its Model Hub and deep link integration with Hugging Face. This update allows users to more easily deploy pretrained models, run performance tests, and manage endpoints for model evaluation.
Amazon SageMaker AI has implemented kernel optimizations for Mixture-of-Experts (MoE) models, specifically focusing on fused MoE operations. These optimizations address inefficiencies in MoE operations, which are critical for scaling large language models while maintaining low per-token compute.
AWS now supports deploying decision models using llama.cpp on Amazon SageMaker AI, offering a managed path for serving quantized models on CPU or GPU. This integration allows for both OpenAI-compatible chat completions and structured decision-making with SemIf, which returns probabilities over fixed choices.
Amazon SageMaker has released the aws-ai-ml skill for the Agent Toolkit for AWS, providing coding agents with expertise in inference optimization and benchmarking. This skill allows agents to generate SageMaker Python SDK v3 code for deploying and evaluating models, bridging the gap between engineering intent and infrastructure specifics.
Amazon SageMaker AI now supports multi-turn reinforcement learning (MTRL) for fine-tuning large language models (LLMs) to create search agents. This method optimizes agent behavior across full multi-turn interactions, offering an alternative to costly expert demonstrations or single-turn optimization.
Amazon Payments implemented a multi-objective contextual multi-armed bandit (MAB) on Amazon SageMaker AI to personalize content in its product acquisition funnel. An A/B test showed a high single-digit percentage relative lift in final-funnel conversion for one customer segment, demonstrating the application of MABs for content selection in the generative AI era.
AWS published a tutorial on deploying real-time voice applications using the vLLM-Omni Deep Learning Container (DLC) on Amazon SageMaker AI. This allows text-to-speech models like Qwen3-TTS to stream audio responses before full generation, enabling interactive voice agents.
AWS published a guide on using vLLM-Omni Deep Learning Containers on SageMaker AI to generate images and videos from text prompts. This process involves deploying separate endpoints for image generation (FLUX.2-klein-4B) and video generation (Wan2.1-VACE-1.3B) from a single container image. The guide demonstrates a workflow where a text prompt creates an image, which is then animated into a video using a motion prompt.
The Qwen3-TTS-12Hz-1.7B-Base text-to-speech model, which supports voice cloning, is now available for deployment on Amazon SageMaker JumpStart. This allows users to generate personalized speech in a target speaker's voice from a short audio recording without model retraining, offering a self-hosted solution for real-time inference.
Amazon SageMaker AI now includes concurrency sweeps, a systematic benchmarking approach to optimize generative AI endpoint configurations. This feature helps users determine the ideal instance type and serving setup to maximize price-performance while maintaining acceptable latency, preventing over-provisioning or under-provisioning of resources.
Positron, Posit's integrated development environment (IDE) for data science, now runs on Amazon SageMaker AI. This integration allows data scientists to use Positron within SageMaker AI for governed data access, R and Python development, model deployment, and reporting, consolidating various data science tasks into a single environment.
Amazon SageMaker Inference has released 13 new capabilities in 2026 across its managed endpoints and HyperPod Inference offerings. These updates aim to simplify generative AI model deployment and optimization for enterprises, startups, and the public sector.
New agent skills allow for the deployment of Hugging Face models on Amazon SageMaker AI, automating the configuration of serving containers, autoscaling, and monitoring. This development addresses challenges with unguided coding agents making incorrect deployment decisions for newer models.
A new method using synthetic data augmentation on Amazon SageMaker AI and Amazon Rekognition generates photorealistic training images for industrial safety AI. This approach addresses the scarcity of real-world data for dangerous scenarios, improving person detection in industrial settings by up to 160% without manual annotation or hazardous data collection. It matters because it provides a safer and more scalable way to train AI models for critical industrial safety applications, particularly for autonomous equipment in sectors like agriculture, construction, and manufacturing.
This guide details how to build an AI-powered product tagging system using Amazon SageMaker's serverless model customization feature. It outlines the process of fine-tuning a smaller open-weight model like Qwen3-8B for consistent and schema-specific product tagging, which is more efficient for high-volume workflows than general-purpose models. This approach matters for businesses needing to automate and standardize product catalog attributes for search, recommendations, and navigation.
Amazon SageMaker now supports instance preference lists for AI training and processing jobs, allowing users to specify an ordered list of up to five acceptable instance types. This feature automatically finds available capacity from the provided list, reducing wait times for GPU resources and improving job start times.
Amazon SageMaker Feature Store now includes an UpdateRecord API, allowing users to modify individual feature values without needing to read and rewrite an entire record. This change reduces latency, prevents race conditions, and lowers costs associated with full-record updates in machine learning feature groups.
This article, Part 2 of a series, details how to implement cross-account model governance using MLflow and Amazon SageMaker AI Model Registry synchronization. It introduces hub-and-spoke and hybrid topologies for managing AI models across different AWS accounts, addressing the needs of larger organizations and regulated environments. This matters for organizations seeking to establish robust, scalable, and secure AI model governance practices across their development and production environments.
Amazon SageMaker AI Model Registry now provides enhanced synchronization with MLflow, automatically transferring training metrics, evaluation results, and lineage. This update allows governance officers to validate and approve models directly from SageMaker, and enables data scientists to manage model lifecycle stages within MLflow.
GitHub has launched Project HydraFusion, a research preview that orchestrates multiple AI models from various providers to complete development tasks. This system dynamically selects execution patterns to balance performance, cost, and latency, aiming to deliver frontier-level quality and cost savings for developers.
ZS Associates developed a security-hardened Amazon SageMaker environment to provide agile ad-hoc analytics while meeting strict compliance requirements for regulated industries. This implementation allows over 1,000 daily active users across more than 200 SageMaker domains to access machine learning capabilities securely.
Amazon SageMaker Feature Store has introduced two new APIs, BatchWriteRecord and ListRecords, to improve data ingestion and discovery capabilities. These additions address previous limitations in handling high-throughput feature pipelines and recovering records from the In-Memory storage tier, making the feature store more efficient and robust for machine learning workflows.
Salesforce implemented Multi-Availability Zone (AZ) high availability for its SageMaker Inference Components (ICs) by utilizing the new SchedulingConfig parameter in the CreateInferenceComponent API. This change addresses a previous limitation where default IC placement, while cost-effective, did not guarantee the necessary AZ distribution for compliance and resilience, particularly for Salesforce's Agentforce AI foundation.
Deepgram has introduced Enhanced Metrics and Prometheus/OpenTelemetry support for its speech AI models deployed on Amazon SageMaker. These additions provide users with detailed usage, billing, and engine-level performance data directly within their AWS accounts, addressing previous observability limitations for self-hosted AI services.
Amazon SageMaker Python SDK v3 updates its script mode, simplifying the process for users to bring their own models for training and deployment without needing to build or maintain Docker images. This update introduces unified classes, ModelTrainer and ModelBuilder, replacing framework-specific estimators, and allows for faster iteration and full container control.
This article, the first in a three-part series, details how to set up a Snowflake environment for a no-code machine learning workflow. It focuses on enabling business users to build predictive models and generate insights without extensive data science or engineering support, addressing challenges faced by organizations with large datasets but limited ML capacity.
This tutorial, the second part of a series, details how to prepare data and build a fraud detection model using Amazon SageMaker Canvas, integrated with Snowflake data sources. It demonstrates visual data transformations with Data Wrangler and model building with XGBoost, enabling business analysts to create ML models without code.
This article, part three of a series, details how to integrate Amazon SageMaker Canvas machine learning predictions with Amazon QuickSight to create interactive dashboards for fraud detection business intelligence. It covers importing Canvas predictions into QuickSight, building analysis dashboards, and using generative BI features for natural language insights. This integration provides a direct path from ML predictions to business-ready dashboards without requiring additional infrastructure.
Amazon introduced the SageMaker AI Spaces add-on for Amazon EKS, allowing data scientists to run interactive IDEs like JupyterLab and Code Editor directly on their EKS clusters. This integration eliminates the need to move workloads off-cluster, providing access to GPU nodes, shared storage, and IAM roles, and can improve GPU utilization by up to 30%.
Amazon SageMaker Python SDK v3 now includes generative AI inference recommendations directly within notebook workflows, automating the benchmarking and deployment optimization of large language models. This integration allows developers to benchmark endpoints, generate data-driven deployment recommendations, and deploy optimized configurations without leaving their notebooks, streamlining the process of optimizing generative AI inference.
AWS released a new solution for inference meta-monitoring on Amazon SageMaker AI endpoints, integrating Amazon Quick to track prediction and data quality metrics. This system provides continuous feedback on model performance in production, addressing silent degradation and data drift issues. It helps ML teams maintain consistent model performance and customer trust by enabling early detection of problems.
Deepgram has integrated AWS IAM Temporary Delegation to improve support for its speech AI models deployed on Amazon SageMaker AI. This integration allows for time-limited, scoped access to customer AWS resources, reducing the time required for initial support investigations from days to minutes.
Amazon SageMaker AI launched a UI for generating inference recommendations, aimed to simplify model deployment. This enables users to obtain optimized configurations quickly, reducing the setup time from hours to minutes without the need for coding expertise.
Amazon SageMaker AI now supports monitoring of discriminative machine learning models using MLflow. This integration allows users to track data and model drift, helping organizations maintain model accuracy amidst changing external factors.
Amazon SageMaker AI now integrates with MLflow, allowing teams to stream benchmark and recommendation results in real-time. This integration streamlines data tracking, reduces silos, and enhances reproducibility in AI inference workflows.