EphAsad/Midas-FableAgent-8B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EphAsad/Midas-FableAgent-8B is an 8.22 billion parameter language model, sequentially fine-tuned from Atem-8B for agentic task execution. It specializes in multi-step task execution, structured action emission, and observation-grounded iteration, operating in execution loops. The model is designed to plan before acting and execute through observation, producing structured JSON action blocks or planning traces. It is optimized for command-line tasks and open-ended planning problems, with a context length of 12,288 tokens.

Loading preview...

Midas-FableAgent-8B: Agentic Task Execution Model

Midas-FableAgent is an 8.22 billion parameter model developed by EphAsad, sequentially fine-tuned from Atem-8B. Unlike its general-purpose reasoning base, Midas-FableAgent is specialized for agentic task execution, designed to operate in iterative loops of planning, acting, observing, and completing tasks.

Key Capabilities

  • Multi-step Task Execution: Excels at breaking down complex goals into concrete actions and iterating based on observations.
  • Structured Action Emission: Produces responses in a defined JSON format for agentic execution, including analysis, plan, commands, and task_complete fields.
  • Observation-Grounded Iteration: Reasons about current states and environment observations to inform subsequent actions.
  • Planning and CoT Reasoning: Capable of generating detailed <think> traces for high-level task decomposition and open-ended planning problems.
  • Sequential Fine-tuning: Utilizes attention-only LoRA to preserve Atem-8B's general reasoning while specializing in agentic formats.

Training Details

The model was trained on two data streams: 10,000 multi-turn agentic execution trajectories from OpenThoughts-Agent-v1-SFT and 4,665 planning and CoT reasoning examples from kelexine/fable-5-sft-traces. It has a maximum sequence length of 12,288 tokens.

Good For

  • Automating command-line tasks in Linux environments.
  • Developing AI agents requiring structured output and iterative execution.
  • Applications needing detailed planning and Chain-of-Thought reasoning for complex problems.

Finetunes
1 models