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

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 19, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

cybertruck32489/LFM2.5-1.2B-Thinking-Fable5-hermes-opus-Agent is a 1.2 billion parameter Hybrid State-Space / Transformer (Liquid Foundation Model) developed by cybertruck32489. This fully fine-tuned model is optimized for multi-domain agentic distillation, combining logical depth with precise API orchestration. It features strict Chain of Thought reasoning and robust parallel tool calling capabilities, making it highly effective for software engineering tasks and advanced mathematical reasoning within a 32,768 token context window.

Loading preview...

Model Overview

This model, LFM2.5-1.2B-Thinking-Fable5-hermes-opus-Agent, is a 1.2 billion parameter Hybrid State-Space / Transformer (Liquid Foundation Model) developed by cybertruck32489. It is a fully fine-tuned, highly optimized agentic distillation designed to integrate the logical reasoning of Claude 3.5 Sonnet/Opus with the precise API orchestration of Hermes into a compact, fast architecture. Continually fine-tuned from LFM2.5-1.2B-Thinking-Fable5-Agent, it excels in software engineering and parallel tool-use scenarios.

Key Capabilities

  • Strict Chain of Thought (CoT): Generates structured, step-by-step task decompositions within <think>...</think> tags before executing actions.
  • Parallel Tool Calling: Supports simultaneous execution of multiple independent functions using LFM bracket notation ([func1(args), func2(args)]).
  • Robust Conversational Balance: Handles standard text-only answers and general reasoning without unnecessary tool invocation.
  • Long Context: Processes inputs up to 32,768 tokens.

Training and Data

The model was trained on 10,756 high-signal step-by-step examples, including Claude Fable-5 traces for software engineering, NousResearch/Hermes Function Calling for multi-turn API trajectories, and Claude Opus Reasoning for advanced mathematics, physics, and logic. It underwent Full Fine-Tuning (FFT) of all 1.17B parameters over 2 epochs, achieving a final training loss of 1.5703 and a best validation loss of 1.6538.

Finetunes
2 models