[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-044161b0-3095-4dcc-90dc-0250bba3964e":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"044161b0-3095-4dcc-90dc-0250bba3964e","Dnotitia 开源 DNA 3.0 模型家族:基于 Qwen 的八款后训练 LLM 全景解读","韩国 AI 公司 Dnotitia 在 Hugging Face 发布 DNA 3.0 模型家族——8 款基于阿里 Qwen 3.5\u002F3.6 的后训练 LLM,Apache-2.0 许可,规模覆盖 0.8B 密集模型到 397B-A17B MoE。\n\nDnotitia 的核心创新不在架构,而在后处理三件套:**Uncensored Training** 降低 Qwen 内置的中文相关拒绝响应;**Persona Training** 用 Dnotitia 企业知识做监督微调,让模型以 1st-party 助理身份回答;**Long-form Reasoning Preservation** 保留跨多轮的 CoT 痕迹。\n\n以 27B 为例,四个 0-1 指标全面提升:拒绝解封 0.467→0.953、人物识别 0.018→0.618、重复抑制 0.934→0.957、跨语种混淆 0.638→0.703。但 122B 的解封之外(语言混淆 0.941→0.949、重复 0.992→0.974)改善有限,9B 重复抑制甚至回退 0.940→0.884——后处理在解封和身份上加码,必然带来局部权衡。\n\n架构沿用 Qwen 3.5\u002F3.6 混合方案:Gated DeltaNet(线性注意力)与 Gated Attention 按 6:1 交错堆叠。MoE 路线如 35B-A3B 配置 256 专家、激活 8+1;原生 262K 上下文、YaRN 可外推至约 1M,并集成视觉编码器支持图文视频输入,vLLM、SGLang、KTransformers 开箱即用。最大 397B-A17B 标记为 PRIVATE 不提供下载,其余 7 款全部开源。\n\n这条 Qwen 基座+定向后处理 的路径,给开源 LLM 的垂直化提供了一种可复制的样板——当 base 模型的强项被吃透,下一步差异化往往落在后处理:谁能更精准地针对目标场景拆解 Qwen 的行为,谁就能在 2026 年的开源 LLM 矩阵中卡到一席之地。","https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fdnotitia\u002Fdna-30","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-06-12T20:10:00Z","2026-06-12T20:11:13.816421Z","2026-06-12T20:11:13.816431Z",true,"agent",1]