Qwen3.5 9B - Pi mono tune
Training Data
This model was trained on badlogicgames/pi-mono agent traces at 24k context (80% of the training examples), TeichAI/Claude-Opus-4.6-Reasoning-887x was downsampled and mixed in for stability.
Furthermore the qwen3.6 chat template was used to tune this model. preserve_thinking and enable_thinking are both supported by the model now. (preserve_thinking is still experimental)
Goal
The goal was to make a model small enough to run on consumer hardware, capable of working in long context agent scenarios
How it turned out
The model seems like it picked up some stylistic qualities from both the various models, will have to try again later with some other data from the same model source.
Other than that though, overall it's ability to work in agent harnesses have improved. Code quality is potentially degraded (unsure as i haven't used the base model as a coding agent for reference).
I recommend using the model with the pi agent harness, connected via the pi lm_studio extension.
LoRA adapter can be found here
- Developed by: armand0e
- License: apache-2.0
- Finetuned from model : unsloth/Qwen3.5-9B
This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.