carlosmm26/Atanor-4B

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Jun 1, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

Atanor-4B is a 4.5 billion parameter model, fine-tuned from Qwen3.5-4B by carlosmm26, specifically optimized for agentic tool-use within the Hermes Agent framework. This model demonstrates improved tool selection and task success rates compared to its base model, making it suitable for local, resource-constrained agentic applications. It was trained entirely on a single RTX 3090, focusing on reasoning repair and Hermes tool-use traces.

Loading preview...

Atanor-4B: Agentic Tool-Use Model

Atanor-4B is a 4.5 billion parameter model, fine-tuned from Qwen3.5-4B by carlosmm26, with a specific focus on enhancing agentic tool-use capabilities within the Hermes Agent ecosystem. This model was developed to explore the potential of smaller models for agentic tasks, with its entire training process conducted locally on a single RTX 3090 GPU.

Key Capabilities & Performance

Evaluated on a 60-task Hermes-native agent benchmark, Atanor-4B shows notable improvements over its base model:

  • Agent Score: Increased from 0.81 to 0.84.
  • Tool Selection: The ability to pick the correct tool for a task doubled from 30% to 60%.
  • Task Success: Improved from 67% to 73%.

Training Methodology

The fine-tuning process involved two LoRA stages (BF16) on an RTX 3090:

  • Stage A (Reasoning Repair): Utilized Bespoke-Stratos and NuminaMath-CoT datasets.
  • Stage B (Hermes Tool-Use): Trained on the kai-os/carnice-glm5-hermes-traces dataset, focusing on agentic traces with a sequence length of 16384.

Good For

  • Local Agentic Applications: Designed for efficient execution in environments like llama.cpp or Hermes Agent, even on consumer-grade GPUs.
  • Tool-Use Scenarios: Excels in tasks requiring precise tool selection and execution.
  • Experimentation: Ideal for developers interested in exploring and deploying smaller, specialized agent models.