| # pipe-agent-v6 | |
| Screenpipe pipe agent v6 — fine-tuned Qwen3.5-9B with **vision enabled**. | |
| - Base: Qwen/Qwen3.5-9B (natively multimodal) | |
| - Training: LoRA r=64, alpha=128, 3 epochs, 7.8K examples | |
| - Eval loss: 0.3210 | |
| - Vision tower frozen during training (pretrained weights preserved) | |
| - See [GGUF version](https://huggingface.co/screenpipe/pipe-agent-v6-GGUF) | |