TeichAI/Qwen3.5-9B-Fable-5-v1

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

TeichAI/Qwen3.5-9B-Fable-5-v1 is a 9 billion parameter language model based on the Qwen3.5 architecture, developed by TeichAI. This model demonstrates improved performance in reasoning tasks, specifically outperforming the base Qwen3.5-9B model on ARC and ARC-E benchmarks. It is optimized for general language understanding and generation, with a notable context length of 32768 tokens. The model is suitable for applications requiring enhanced logical inference capabilities.

Loading preview...

TeichAI/Qwen3.5-9B-Fable-5-v1 Overview

TeichAI/Qwen3.5-9B-Fable-5-v1 is a 9 billion parameter language model built upon the Qwen3.5 architecture. This version represents a fine-tuned iteration that shows enhanced performance compared to its base model, Qwen3.5-9B, particularly in reasoning-focused benchmarks.

Key Capabilities and Performance

This model demonstrates notable improvements in specific reasoning tasks. Benchmarking against the base Qwen3.5-9B model reveals:

  • ARC (AI2 Reasoning Challenge): Achieves 0.624, an increase from Qwen3.5-9B's 0.553.
  • ARC-E (ARC Easy): Scores 0.806, up from Qwen3.5-9B's 0.712.
  • BoolQ: Maintains strong performance at 0.891, comparable to the base model's 0.892.

These results suggest a more successful optimization for tasks requiring logical inference and understanding. The model also benefits from a substantial context length of 32768 tokens, allowing for processing longer inputs and generating more coherent, extended outputs.

Training and Development

The training process for this Qwen3.5 model leveraged Teich for data extraction, formatting, and masking. Additionally, the training was accelerated using Unsloth and Huggingface's TRL library, achieving a 2x faster training speed. This indicates an efficient and streamlined development approach.

Good For

  • Applications requiring improved reasoning and logical inference capabilities.
  • Tasks benefiting from a large context window (32768 tokens).
  • General language generation and understanding where enhanced performance over base Qwen3.5-9B is desired.