This model is a merged version of deepreinforce-ai/Ornith-1.0-9B and the MTP (Multi-Token Prediction) head provided by protoLabsAI.

The original Ornith-1.0-9B model was shipped without the mtp.* tensors present in its Qwen3.5-9B base, preventing native speculative decoding speedups. This repository provides the re-aligned MTP head, specifically retrained to match Ornith's hidden states for lossless throughput acceleration.