Qwen3.5_2B-Fable-trained

Qwen3.5-2B Fable Agent (GGUF, Q4_K_M)

A personal fine-tuning experiment: Qwen3.5-2B fine-tuned on the agent and reasoning traces in Glint-Research/Fable-5-traces, then quantized to a 4-bit GGUF file for local inference with llama.cpp-compatible runtimes.

Model details

• Developed by: Shiheng Wang (personal project; the training scripts were written with help from AI coding tools)
• Base model: Qwen/Qwen3.5-2B
• Training data: Glint-Research/Fable-5-traces
• File: qwen35-2b-agent-q4_k_m.gguf (Q4_K_M quantization, about 1.3 GB)
• Parameters: about 1.94 B (from the GGUF metadata)
• Context length in the GGUF metadata: 262,144 tokens
• Language: mainly English
• License: Apache-2.0. Please also check the licences of the base model and the dataset.
• Released: July 2026

How to run

With llama.cpp:

huggingface-cli download wshhghghgh/Qwen3.5_2B-Fable-trained qwen35-2b-agent-q4_k_m.gguf --local-dir .
llama-cli -m qwen35-2b-agent-q4_k_m.gguf -cnv

With Ollama:

ollama run hf.co/wshhghghgh/Qwen3.5_2B-Fable-trained

Training

The base model was fine-tuned on the dataset's agent trajectories and reasoning traces, using the Qwen chat template, and the result was exported to GGUF and quantized to Q4_K_M. Detailed training settings (method, steps, hardware) were not recorded for this release.

Evaluation

No benchmark evaluation has been run. Treat this model as an experiment, not as a tuned or validated release.

Limitations

• It is a small model at 4-bit precision, so it can produce wrong facts, flawed reasoning and broken code that still looks convincing.
• Check its output before relying on it, especially code.
• Do not use it for medical, legal, financial or other safety-critical decisions.
• It inherits the limitations and biases of the base model and the training data.

Credits

• Base model: the Qwen team (Qwen/Qwen3.5-2B)
• Dataset: Glint Research (Glint-Research/Fable-5-traces)