Tencent has released a new speculative decoding method called DSpine that reportedly achieves 3x faster decoding speeds for 27B parameter models. The implementation is available on Hugging Face under the model identifier Qwen3.8-27B-DSpine-b16.