atenareply/gemma-4-12b-asterion-agentic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The atenareply/gemma-4-12b-asterion-agentic is a 12 billion parameter Gemma-4 based language model fine-tuned for agentic behavior in a verifiable maintenance assistant role. It specializes in telemetry anomaly triage, native Gemma-4 tool-calling, and raw-log reading within a fictional satellite operations domain. This model is optimized for integration into tool-execution loops, enabling it to classify alerts and manage maintenance tasks for a 24-satellite constellation. Its training incorporates a unique native-template offset masking and a dataset designed to fix confounds in raw-vs-clean tool output.

Loading preview...

Overview

The atenareply/gemma-4-12b-asterion-agentic is a 12 billion parameter model built upon the Gemma-4 architecture, specifically fine-tuned for agentic capabilities. It functions as a verifiable maintenance assistant, primarily designed for telemetry anomaly triage, native Gemma-4 tool-calling, and processing raw log data within a simulated satellite operations environment. The model's lineage traces from gemma-4-12B through CPT and ParamΔ stages, culminating in this Agentic SFT (Supervised Fine-Tuning) version.

Key Capabilities

  • Agentic Tool-Calling: Utilizes native Gemma-4 tool-calling for interacting with 8 predefined tools, emitting <|tool_call> blocks and consuming <|tool_response> results.
  • Telemetry Anomaly Triage: Specializes in classifying alerts (routine vs. genuine) and initiating maintenance loops for genuine anomalies in a 24-satellite constellation scenario.
  • Raw-Log Reading: Capable of interpreting and acting upon raw telemetry logs.
  • High Extraction Accuracy: Achieves 0.90 pass@1 for extraction tasks, with 0.96 per-question accuracy.
  • Robust Tool Integration: Demonstrates perfect schema, grounding, and budget adherence (1.00 for each) with its wire format.

Good For

  • Automated Maintenance Systems: Ideal for use cases requiring an AI agent to classify system alerts and manage operational workflows within a tool-execution loop.
  • Simulated Operational Environments: Particularly suited for applications in fictional or simulated domains like satellite operations, where verifiable agentic behavior is critical.
  • Research in Agentic LLMs: Offers insights into advanced SFT methods, including native-template offset masking and confound-fixed datasets for tool-grounded trajectories.

Limitations

  • Requires a multi-turn tool-execution loop to function correctly.
  • The domain is fictional, and the 'genuine' alert class is largely synthetic, meaning evaluation measures verifier-conformance rather than real-world detection.