Ammonix/AmmonixWtE-Writer-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Ammonix/AmmonixWtE-Writer-9B is a 9 billion parameter language model developed by Ammonix, specifically designed to write action payloads for industrial control-room agents. This LoRA-trained model, with a 32768 token context length, excels at generating structured outputs like bunker assignments or blend ratios within a simulated waste-to-energy plant environment. It is optimized for writing specific, validated actions rather than making decisions, reducing hard-limit violations in simulated industrial control scenarios.

Loading preview...

AmmonixWtE-Writer-9B: Industrial Control-Room Action Payload Writer

AmmonixWtE-Writer-9B is a specialized 9 billion parameter language model developed by Ammonix, serving as the frozen local language model component of the Ammonix Industrial Control-Room Agent. Its primary function is to write action payloads (e.g., bunker assignments, blend ratios, load targets) after a decision has been made by the agent, rather than making the decisions itself. All generated payloads must pass a runtime validator before execution.

Key Capabilities & Features

  • Specialized Payload Generation: Designed exclusively for generating structured action payloads within an industrial control context.
  • Outcome-Selected SFT: Trained using a unique outcome-selected Supervised Fine-Tuning (SFT) method on 160 fully synthetic payload pairs, without a teacher model.
  • Reduced Hard-Limit Violations: In simulated evaluations, this model demonstrated a reduction in hard-limit violations compared to the base model, even if overall success rates were statistically similar.
  • LoRA Fine-Tuned: A LoRA (r=32) adapter was trained for 2 epochs, taking approximately 48 minutes on a consumer GPU.
  • Synthetic Data Training: Entirely trained and evaluated on fully synthetic data from a simulated waste-to-energy plant, ensuring no real-world plant, operator, or supplier data was used.

Good For

  • Industrial Control-Room Automation: Ideal for applications requiring precise, validated action payload generation within a simulated industrial control environment.
  • Research and Evaluation: Suitable for research, educational, and evaluation purposes related to industrial control systems and agent-based automation.
  • Structured Output Generation: Excels at producing structured JSON outputs that adhere to predefined schemas for specific industrial actions.

Limitations

It is crucial to note that this model was trained and evaluated exclusively on a synthetic plant and has never interacted with a real facility. It functions as a writer, not a decision-maker or safety function, and its outputs are unchecked outside the agent's harness. It is not intended for operating a real facility.