Skip to main content
Newton is a foundation model for reasoning about and acting on the physical world. You can customize Newton in three primary ways, progressing from lightweight configuration to deeper adaptation using your own data. Not every method applies to every model family: Newton Fusion is steered with natural-language prompts, while Newton Omega is a time-series encoder that takes no prompt — its behavior is shaped entirely by the data you give it. These customization options work together, letting you apply one or several to tailor Newton to your specific needs. Each approach varies in complexity and cost, so we suggest beginning with the lightest method available for your model and advancing only as your requirements evolve.

​
Prompting

Applies to: Newton Fusion. The Fusion model is steered at query time through the Direct Query API — no training, no configuration to manage:
  • Instructions — set the model’s role and behavior with instruction_prompt (for example, “You are a perception agent analyzing activity in front of a house. Reply in 2-3 sentences.”).
  • Output format — make the prompt the schema: ask for “ONLY a JSON object: {"activity": "...", "confidence": <0..1>}” and parse the response directly.
  • Generation parameters — control length and sampling with max_new_tokens, temperature, and related parameters.
The Developer Quickstart shows prompting on a live video example. Omega has no prompting tier: it receives raw sensor windows and returns embeddings, so customization starts at level 2.

​
N-Shot Examples

Applies to: Newton Fusion & Newton Omega — through different mechanisms. A small set of labeled examples steers the model without any training:
  • Fusion — include labeled input/output examples in the prompt itself so the model imitates them, the same way you set instructions and output format in Prompting.
  • Omega — build a labeled reference library of embedded sensor windows and classify new windows against it with a nearest-neighbor vote. See N-Shot Examples; the quickstart’s fault-detection example reaches 95% accuracy from 8 labeled windows this way.

​
Fine-Tuning

Applies to: Newton Fusion & Newton Omega. When you have hundreds of labeled examples per class and need maximum accuracy, the Fine-Tuning Service trains a dedicated checkpoint on your dataset. Fine-tuned checkpoints are selected by model ID at query time, so they drop into existing Direct Query and Batch workflows.

​
Choosing the Right Approach

Start simple: Begin with prompting (Fusion) or a small n-shot reference library (Omega), and consider fine-tuning only when you have sufficient data and need maximum accuracy. Choosing the right customization method based on your data: