https://pillumina.github.io/tags/ornith/ Recent content in Ornith on CctoctoFX https://pillumina.github.io/imgs/icon_head.png https://pillumina.github.io/imgs/icon_head.png Hugo -- 0.148.2 en Mon, 29 Jun 2026 00:00:00 +0000 https://pillumina.github.io/posts/qa/ornith-1.0-397b-qa/ Mon, 29 Jun 2026 00:00:00 +0000 https://pillumina.github.io/posts/qa/ornith-1.0-397b-qa/ Ornith-1.0-397B 架构 50 问，覆盖总参/激活/KV Cache、hybrid attention 3:1 设计、512E top-10 MoE、self-scaffolding RL、reward hacking 防御、位置编码及推理部署。 https://pillumina.github.io/posts/architecture/ornith-1.0-397b/ Mon, 29 Jun 2026 00:00:00 +0000 https://pillumina.github.io/posts/architecture/ornith-1.0-397b/ Ornith-1.0-397B 是 DeepReinforce AI 的 frontier-scale agentic coding 模型。核心创新在 self-scaffolding RL 后训练方法，架构层面继承 Qwen3.5-MoE：60层 hybrid attention（45线性+15 full）、512路 top-10 MoE FFN、1层 MTP 辅助头、256K 原生上下文、27层 ViT。本期完整拆解 397B 规模配置、FLOPs/KV Cache/推理显存、self-scaffolding RL 框架、reward hacking 防御及异步 GRPO 训练体系。