The Pulse
OpenTSLM time-series model becomes available in SageMaker JumpStart
AWS customers can deploy OpenTSLM-Flamingo to a SageMaker AI inference endpoint in their own infrastructure. The model combines numerical time series with text context, and AWS describes an illustrative IT-support workflow using it.

OpenTSLM-Flamingo is now available through Amazon SageMaker JumpStart, giving AWS customers a way to deploy the time-series language model in their own cloud environment. The model can be deployed to a SageMaker AI inference endpoint and used to combine numerical measurements with text context and questions, according to an AWS post published Nov. 17.
Aionic Labs is developing OpenTSLM as an open-source family of time-series language models. The models are designed to interpret measurements such as equipment sensor readings or supply and demand alongside material including maintenance records, market news and engineers’ observations.
How the model is deployed
The JumpStart model combines the Gemma 3 4B language model with a time-series encoder and cross-attention layers that allow the language model to attend to the encoded series. It has 5.24 billion parameters stored in bfloat16. AWS says its weights take about 10.5 GB of GPU memory.
An application sends numerical time series and text context to the SageMaker endpoint and receives a response. The application can present that response to an expert or use it in another workflow. It can also select the observation window and signals and validate the response before using it.
The model runs on a single GPU with at least 24 GB of memory and bfloat16 support; AWS recommends 32 GB of host memory. The default instance type is ml.g6.2xlarge, with one NVIDIA L4 GPU, 24 GB of GPU memory and 32 GB of host memory. Other supported instance types are listed in the walkthrough.
Deployment through SageMaker Studio requires users to review the Gemma Terms of Use and accept the end-user license agreement. The deployment walkthrough also requires an AWS account with a SageMaker AI domain and an execution role able to create models and endpoints in a supported region. For production, AWS recommends a policy that grants only the permissions needed.
An IT-support example
Nexthink and Aionic Labs are working on proactive IT assistance using Nexthink’s personal IT agent, Spark, with time-series language models. Their workflow combines device measurements, including CPU activity, memory use, disk space, application stability and connectivity, with timestamped events and device and software context.
AWS describes the model as providing Spark with a state summary, a recommended action and an expected outcome. The summary connects the likely issue and signal patterns to the employee’s experience, while the recommended action specifies checks and a practical action for the agent to offer.
In an illustrative low-disk-space scenario, the model connects falling free space, a rising disk queue and more frequent application hangs with possible symptoms such as a slow computer. Spark checks available space and whether an IT-approved cleanup is appropriate; if the checks confirm it is, Spark asks the employee to confirm before running it. The example includes a follow-up check and an offer to pass findings to IT support if the issue persists.
Fine-tuning walkthrough
The AWS walkthrough also describes fine-tuning a pretrained OpenTSLM checkpoint from Hugging Face using ARFBench data loaded through TimeNet. TimeNet provides a format and API for finding, building and loading datasets used to train and evaluate the models.
The notebook submits a SageMaker training job that loads the data and fine-tunes the model. Amazon S3 stores the resulting model artifact, which is deployed to a separate real-time inference endpoint for evaluation. The fine-tuning example uses an ml.g7e.2xlarge instance; the notebook itself does not need a GPU.