LoginSign Up
← Back to all sparks
A

AWS Machine Learning

AI-ASSISTANTS
Velocity10.0

Amazon Web Services' official AI/ML blog covering Bedrock, SageMaker, AgentCore, and Nova model updates.

aws.amazon.com/blogs/machine-learning ↗

AWS keeps widening Bedrock's model catalog and deepening Nova and agent infra

awsbedrockamazon-novasagemakermodel-integrationsagents
◆Current state
The AWS Machine Learning feed is a high-frequency stream of model integrations, Nova capabilities, and solution walkthroughs. This period covers a one-click Hugging Face-to-SageMaker Studio deep link, MiniMax models arriving on Bedrock, a Nova selective-unlearning technique behind Customizable Content Moderation, multi-turn RL infrastructure on SageMaker HyperPod, a Nova-directed PII-redaction pipeline, and MLflow streaming for SageMaker benchmarks. Individually incremental, collectively a steady platform build-out.
◆Where it's heading
Two consistent vectors: Bedrock as a model-agnostic hub (MiniMax now, GPT-OSS and Nemotron in GovCloud just outside this window) and Nova as AWS's first-party family gaining moderation, vision, and unlearning capabilities. Layered on top is agentic and RL infrastructure — HyperPod multi-turn RL, a serverless A2A gateway for agent routing. AWS is positioning SageMaker and Bedrock as the operational substrate for both third-party and first-party models plus the agents built on them.
◆Prediction
Expect continued model-catalog additions to Bedrock and further Nova capability and agent-infrastructure posts. The through-line — reducing friction from model discovery to training to agent deployment on AWS — is the safe bet for the next batch.

◆Recent moves

  1. 17h ago

    From Hugging Face to Amazon SageMaker Studio in one click

    A new deep-link integration takes developers from a Hugging Face model straight into SageMaker Studio in one selection. A friction-reducing tie between the two ecosystems, incremental but real.

    View source ↗
  2. 17h ago

    Teaching models to forget: Selective unlearning with Amazon Nova

    Introduces rDPO, a reverse-preference-optimization unlearning technique behind Amazon Nova's Customizable Content Moderation, aimed at cutting over-deflection while preserving quality. A real Nova capability plus a reusable method.

    View source ↗
  3. 22h ago

    Run MiniMax models on Amazon Bedrock

    MiniMax models are now available on Amazon Bedrock with on-demand scaling and multiple APIs, extending Bedrock's third-party catalog for agentic and long-context workloads.

    View source ↗
  4. 22h ago

    Deploying Multi-Turn RL Infrastructure for Amazon Nova on Amazon SageMaker HyperPod

    A walkthrough for deploying event-driven, two-phase multi-turn RL for Amazon Nova via Nova Forge on SageMaker HyperPod. Reference infrastructure that lowers the bar for RL training pipelines on AWS.

    View source ↗
  5. 22h ago

    Automatically redact PII in images with Amazon Nova

    A Nova-directed pipeline coordinates SAM 3 segmentation and Textract OCR to redact PII in images, including hard cases like ID cards and license plates. A compliance-oriented solution built on Nova's vision reasoning.

    View source ↗
  6. 22h ago

    Streaming benchmark and recommendation results to MLflow with Amazon SageMaker AI

    A new MLflow integration streams SageMaker inference-recommendation and benchmark metrics into a unified tracking interface in real time. Incremental tooling that consolidates experiment tracking.

    View source ↗