腾讯混元 · Hugging Face 新模型·· 7 天前22
混元发布 Qwen3.8-27B-DSpine-b16 预训练模型
tencent/Qwen3.8-27B-DSpine-b16
要点速览
混元发布 Qwen3.8-27B-DSpine-b16 预训练模型，包含 27B 参数、b16 精度、5 层 Draft 层和 8,192 token 滑动窗口，用于推理时提升效率，需配合 Qwen3.8-27B 模型使用。
正文
DSpine draft weights for Qwen3.8-27B.
| Property | Value |
|---|---|
| Draft layers | 5 |
| Block size | 16 |
| Sliding window | 8,192 tokens |
| Precision | BF16 |
Training started on October 27, 2026. This checkpoint is from data step 18,210 and optimizer step 18,000.
Usage
Follow the Qwen3.8-27B inference instructions
in DSpine. Set --model-path to Qwen/Qwen3.8-27B and
--speculative-draft-model-path to this model repository or its local directory.
The tokenizer, token embeddings and LM head are supplied by the target model.
来源: 腾讯混元 · Hugging Face 新模型 · huggingface.co