Model grafting: turning Qwen3.5-4B into a causal encoder-decoder after the fact

Recently, the new DeepSeek-V4.1-Flash architecture showed how a causal encoder-decoder can work, but it was trained from scratch. Model Grafting does it to an existing model: cut at some depth, let the lower layers read the prompt, and use the upper layers get for encoder's residual stream as prefix KV via identity-init adapters, then heal with self-distillation from the unmodified parent. Decoding part stays the same, this method was described in this blog post https://latentnode.pages.dev/articles/model-grafting

I applied the same recipe to Qwen3.5-4B to create two graft variants - https://huggingface.co/LocalLLaMA/Qwen3.5-4B-graft8 and https://huggingface.co/LocalLLaMA/Qwen3.5-4B-graft16

The graft8 variant shows speedup of ~3.7x at 128K prompt with some loss in accuracy. The graft16 is much more closer to the parent model while showing 2.0x speedup in prompt processing with minimum loss in accuracy.