Model grafting: turning Qwen3.5-4B into a causal encoder-decoder after the fact
英文标题：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
打开原文相关 2 · 机会 0
中文
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
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
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