cybertruck32489/LFM2.5-1.2B-Thinking-Fable5-Agent

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 19, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

cybertruck32489/LFM2.5-1.2B-Thinking-Fable5-Agent is a 1.2 billion parameter Hybrid State-Space / Transformer (Liquid Foundation Model) fine-tuned for autonomous software engineering and terminal agent tasks. Developed by cybertruck32489, it excels at structured reasoning within tags and precise Python-style tool orchestration for file operations (Read, Write, Edit) and shell execution (Bash). This model, with a 32,768 token context length, is specifically designed to mimic advanced coding agent behaviors.

Loading preview...

LFM 2.5 1.2B Thinking - Fable-5 SE Agent Overview

This model, developed by cybertruck32489, is the first checkpoint of a fully fine-tuned (FFT) Liquid Foundation Model (LFM 2.5) with 1.2 billion parameters. It is specifically optimized to function as an autonomous software engineering and terminal agent, distilling the reasoning and action trajectories of advanced coding agents. The model is based on PinoCookie/LFM2.5-1.2B-Thinking-Abliterated and supports a substantial context length of 32,768 tokens.

Key Capabilities

  • Structured Reasoning: Generates detailed, step-by-step thought processes encapsulated within <think>...</think> tags, mimicking high-signal reasoning.
  • Standardized Tool Use: Emits precise Python-style tool execution syntax, wrapped in special tokens, for reliable automation.
  • Terminal & Code Operations: Fine-tuned for specific actions like Read, Write, Edit for file manipulation, and Bash for shell command execution.

Training and Optimization

The model was trained using Full Fine-Tuning on the Glint-Research/Fable-5-traces dataset. A key aspect of its training involved loss masking, where loss was calculated only on the assistant's responses (thinking blocks and tool calls), preventing the model from simulating terminal outputs or user inputs. This approach ensures the model focuses on generating actionable agent behaviors.

Good for

  • Automated software development tasks.
  • Building intelligent agents that interact with file systems and shell environments.
  • Applications requiring structured reasoning and precise tool orchestration in a coding context.

Finetunes
1 models