AWS Strengthens SageMaker with Native Data and Temporal Model

AWS adds generation, evaluation, and data transformation tools to SageMaker AI, while making a temporal model available via JumpStart. Use cases target both dataset preparation and operational scenarios such as IT support.

On November 17, 2026, AWS announced two new features related to SageMaker AI: built-in data preparation capabilities and the availability of OpenTSLM-Flamingo Gemma 3 4B in JumpStart. The first aims to simplify the creation, validation, and annotation of datasets; the second provides a model capable of interpreting time series with text, useful for cases like maintenance or IT support. The challenge is very concrete: reducing patchwork between multiple tools while maintaining control over data and validation steps.

Des données jusqu’au modèle, sans quitter le même flux

With SageMaker AI, AWS now combines synthetic data generation, quality evaluation, transformation, auto-labeling, and human review into a single workflow. Teams can chain these steps or use only one, with specifications and code inspectable at each stage. This approach targets a well-known problem for data teams: datasets often circulate between homegrown scripts, annotation tools, and separate checks, slowing iterations.

Synthetic generation produces records for training, evaluation, or testing from a specification, optionally fed by seed data and documents. Evaluation can verify business rules, built-in metrics, and LLM-as-a-judge metrics, where a model judges data quality. Then, transformation can deduplicate, merge, filter, or reformat entries via generated code or user-provided Python. Everything remains executable at service scale, avoiding the team having to move CSV files for three days.

Quand l’auto-étiquetage renvoie les cas douteux à l’humain

The auto-labeling component works from an unlabeled set, a small already annotated sample, or even without initial labels. AWS supports tasks like classification, named entity recognition, extractive question answering, ordinal evaluation, and summarization, with the possibility to define custom tasks. For closed-answer problems, multiple models annotate each line and their agreement level serves as a confidence score; cases below the chosen threshold then switch to human review in SageMaker Studio, via Label Studio Community Edition.

Specialist corrections return to the auto-labeling loop, enabling consensus metrics and annotation reports. AWS also specifies that generation, auto-labeling, evaluation, and transformation jobs are managed by SageMaker, while generated data is stored in the client’s S3 buckets, in the chosen AWS region. In practice, automation advances without datasets leaving the perimeter controlled by the client.

OpenTSLM-Flamingo Gemma 3 4B passe en production, mais sous contraintes

In the same move, AWS made OpenTSLM-Flamingo Gemma 3 4B available in SageMaker JumpStart, its hub of ready-to-deploy models. This 5.24 billion parameter model, which combines Gemma 3 4B with a time series encoder and cross-attention layers, can be deployed on a SageMaker AI endpoint in real time within the client’s AWS environment. Its size occupies about 10.5 GB of GPU memory in bfloat16, with a default deployment on an ml.g6.2xlarge instance equipped with a 24 GB NVIDIA L4.

OpenTSLM is designed to interpret numerical measurements with text, such as maintenance logs, market news, or engineer observations, then produce natural language explanations. The Flamingo version uses a Perceiver Resampler to create a fixed-size latent representation, then gated cross-attention to integrate it into the language model. This architecture limits memory usage but does not guarantee that a brief event will survive intact in a very long record; AWS therefore recommends testing representative windows, maintaining sufficient resolution for short events, and indicating units, sampling intervals, and observation periods.

Le support IT comme terrain d’essai

AWS illustrates the use case with Spark, Nexthink’s personal IT agent. The workflow combines recent CPU, memory, disk, application stability, and connectivity measurements, plus timestamped events, device configurations, and software versions. The model then provides a status summary, recommended check, proposed action, and expected result that Spark can verify.

In the example of low disk space, the system links free space decrease, disk queue increase, and application freeze frequency to symptoms like slowdowns or saving difficulties. Spark then checks if an IT-approved cleanup is appropriate, shows what would be deleted, and asks for confirmation before acting. If the problem persists, it can switch to human support. The setting is simple, but the promise is clear: moving from a raw alert to guided action.

Ce que disent les chiffres

  • 45.9% — accuracy after fine-tuning on 135 selected ARFBench questions.
  • 5.0% — accuracy before fine-tuning on the same ARFBench questions.
  • 750 — multiple-choice questions in ARFBench.
  • 142 — time series from 63 real internal incidents.
  • 60 seconds — real-time call limit on the default AWS endpoint.

Pourquoi ce double mouvement compte

These announcements tell the same strategy from two angles. On one side, AWS tries to reduce friction around training data, where quality and iteration speed often hinge. On the other, the platform pushes a temporal model toward operational use, with simpler packaging for teams wanting to process business signals without building their own service stack. For companies, the issue is not only model performance; it is also the share of work that disappears between collection, verification, and production deployment.

The framework remains limited, however. Real-time calls are capped at 60 seconds, request size at 6 MB, and the model relies on precise analysis windows rather than unlimited memory. AWS also recommends accepting the Gemma Terms of Use for deploying the JumpStart model. In other words, the promise is of smoother integration, not a model that understands everything, everywhere, all the time.

To read