Models

AWS Puts OpenTSLM Time-Series Reasoning Model on SageMaker JumpStart

AWS added OpenTSLM-Flamingo Gemma 3 4B to SageMaker JumpStart on November 17, 2026; fine-tuning raised ARFBench incident-question accuracy from 5.0% to 45.9%.

AWS Brings OpenTSLM Time-Series Reasoning to SageMaker JumpStart
AWS Brings OpenTSLM Time-Series Reasoning to SageMaker JumpStartCampaign Creators / unsplash
By Marcus Bennett6 min read

Why it matters

  • AWS made OpenTSLM-Flamingo Gemma 3 4B (5.24 billion parameters) available in SageMaker JumpStart on November 17, 2026.
  • Fine-tuning on ARFBench data raised accuracy on 135 held-out incident questions from 5.0% to 45.9% on one ml.g7e.2xlarge instance.
  • The default ml.g6.2xlarge endpoint runs one 24 GB NVIDIA L4; requests accept up to four instances, 32 series each and 32,768 values per series.
  • Datadog's ARFBench contains 750 question-answer pairs built from 142 time series across 63 real internal incidents.
  • The model weights occupy about 10.5 GB of GPU memory in bfloat16, with real-time calls capped at 60 seconds and 6 MB.

Fine-tuning lifted accuracy on held-out incident questions from 5.0% to 45.9%, and that result now ships inside a managed AWS product. On November 17, 2026, Amazon Web Services announced that OpenTSLM-Flamingo Gemma 3 4B is available in Amazon SageMaker JumpStart, AWS's machine-learning model hub, letting customers deploy a 5.24-billion-parameter model to a real-time inference endpoint inside their own AWS environments.

The release matters because it packages a research-grade architecture — one that treats numerical time series as a native input to a language model — behind a managed deployment path. Teams that operate on telemetry, sensor streams or medical signals can now run the model without assembling the training and serving infrastructure themselves.

What does the model actually do?

OpenTSLM interprets numerical measurements alongside text such as maintenance records, market news and engineers' observations, then returns natural-language explanations and answers. The JumpStart package pairs Gemma 3 4B with a time-series encoder and cross-attention layers, which let the language model attend to the encoded signals. Its weights occupy about 10.5 GB of GPU memory when stored in bfloat16, a numeric format that uses 16 bits per parameter.

The underlying OpenTSLM research paper distinguishes two architectures. Both treat time series as a native input modality — a data type the model processes directly rather than converting to text first — but they integrate it differently:

  • SoftPrompt adds learned time-series tokens to the language model's input sequence.
  • Flamingo, the variant AWS released, uses a Perceiver Resampler (a component that compresses variable-length inputs into a fixed-size latent representation) and gated cross-attention to inject that representation into the model.

The original evaluation covered ECG question answering, sleep staging and human activity recognition, giving the managed release a research foundation rooted in medical and wearable data rather than IT operations.

The separate OpenTSLM repository is MIT-licensed and supplies fine-tuning code, with pretrained checkpoints available through the project's Hugging Face organization. The project identifies itself as a collaboration between Stanford University's Schmiedmayer Lab and ETH Zurich's Agentic Systems Lab.

How does Nexthink use it?

AWS describes Nexthink and Aionic Labs as building toward proactive assistance through Spark, Nexthink's personal IT agent, so employees can receive help before asking for it. The workflow combines recent CPU, memory, disk, application-stability and connectivity measurements with timestamped events, device configuration and software versions. The model then supplies a state summary with supporting signals, a recommended check and action, and an expected outcome that Spark can verify.

AWS's low-disk-space example illustrates the pattern. The model connects falling free space, a rising disk queue and more frequent application hangs to symptoms such as slow performance or difficulty saving files. Spark checks whether an IT-approved cleanup is appropriate, shows the employee what would be removed and seeks confirmation before acting. It then verifies the result and offers escalation to IT support if the problem persists.

That confirmation step grounds the story's stakes: the model recommends, the agent verifies, and a human approves. For enterprise IT, the difference between a suggestion and an automated action determines whether such a workflow passes internal review.

What are the deployment limits?

AWS's default ml.g6.2xlarge endpoint meets the single-GPU, bfloat16 requirement with one 24 GB NVIDIA L4 and 32 GB of host memory. A request can contain up to four question instances, each with one to 32 numerical series and as many as 32,768 values per series. Deploying the gated JumpStart model requires accepting the Gemma Terms of Use.

The endpoint standardizes each series to a mean of zero and a standard deviation of one, while supplying the original mean and standard deviation as text so the model receives both shape and scale. Real-time calls face a 60-second limit and a 6 MB request-body cap. AWS reports roughly 20 generated tokens per second on the default instance. Instances are processed sequentially, and the company recommends asynchronous inference for longer-running work.

Does Flamingo's compression lose information?

Yes, in certain conditions. Flamingo's fixed-size representation bounds what reaches the language model, but it does not give the complete system unlimited memory, and it cannot guarantee that a brief event remains identifiable in a very long recording. AWS advises operational teams to:

  • test representative normal and abnormal windows;
  • retain sampling resolution for short events;
  • align related channels;
  • include units, sampling intervals and observation periods in the context.

The payload cap and the model's useful analysis window are separate constraints, which means a request that fits within 6 MB may still exceed what the architecture can meaningfully analyze.

How does fine-tuning work — and what did it achieve?

AWS's adaptation workflow uses TimeNet, which provides a common dataset format and Python API with connectors and immutable versions. TimeNet's documentation explicitly limits its scope to data; training, inference, model definitions and evaluation metrics fall outside it. The example workflow loads a pretrained checkpoint from Hugging Face and ARFBench data through TimeNet, runs a SageMaker training job, stores the adapted model in Amazon S3 and deploys it to a separate real-time endpoint using the same request format.

Datadog's ARFBench (an incident-response benchmark) contains 750 multiple-choice question-answer pairs derived from 142 time series across 63 real internal incidents, using production telemetry and additional context. The training job updates the time-series encoder, Perceiver Resampler, cross-attention layers and language-model token embeddings while keeping the rest of Gemma 3 frozen.

The headline number came from AWS's own run on one ml.g7e.2xlarge instance. Accuracy on 135 held-out ARFBench questions rose from 5.0% before fine-tuning to 45.9% afterward. The pretrained model left most answers empty, while the adapted model answered every question with a supplied option. AWS states the run demonstrates adaptation to a new task and answer format; it is not a comparison with other models or an official ARFBench leaderboard result.

That caveat carries weight. A 45.9% score on a multiple-choice benchmark shows the model can learn an answer format and internal telemetry distribution, but it does not establish superiority over competing approaches, and no independent leaderboard entry currently validates the result.

What happens next?

The JumpStart release makes a native time-series architecture easier to deploy and adapt, while its evidence remains tied to particular windows, tasks and answer formats. For operational teams, the central question is whether reviewed, representative examples show that the model's explanations and recommended actions are reliable enough for the workflows they are meant to support — a question each adopter will have to answer with its own data before granting Spark permission to act.

Source: unite.ai

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

718 articles

Related articles

  1. StepFun opens Step 5 Preview weights: 600B MoE, 27B active, 1M context
  2. OpenAI Launches o3 and o4-mini, Its Smartest Models Yet
  3. OpenAI Ships o1 to Developers With 60% Cheaper Audio
  4. OpenAI Ships GPT-5.3-Codex-Spark, a Coding Model Built for Real-Time Work
  5. OpenAI frontier models and Codex land on AWS Bedrock

« Previous articleNext article »