diff --git a/honcho/env.cfg.example b/honcho/env.cfg.example
index e17cf0a..7a6c8fb 100644
--- a/honcho/env.cfg.example
+++ b/honcho/env.cfg.example
@@ -61,14 +61,19 @@ HONCHO_EMBEDDING_BASE_URL=https://api.siliconflow.cn/v1
 # 必须与已有 DB schema 的 public.documents.embedding dim 一致；切换模型必须先 migrate 旧数据
 # 若 Honcho 自动推断失败，需在 stack.yml 同步设置 EMBEDDING_VECTOR_DIMENSIONS=1536
 # deriver 必须用支持 tool calling 的模型
-HONCHO_DERIVER_MODEL=gpt-5.4-mini
+HONCHO_DERIVER_MODEL=deepseek-ai/DeepSeek-V4-Flash
+# dreamer 的 deduction/induction specialist（默认走 OpenAI SDK + gpt-5.4-mini，
+# 会请求 api.openai.com/v1 在国内超时）；必须显式覆盖到国内可达端点
+HONCHO_DEDUCTION_MODEL=deepseek-ai/DeepSeek-V4-Flash
+HONCHO_INDUCTION_MODEL=deepseek-ai/DeepSeek-V4-Flash
 HONCHO_DERIVER_WORKERS=1
 
 # ============================================================
 # LLM（OpenAI 兼容协议；可换 OpenRouter / vLLM / Ollama 等端点）
 # deriver 启动时强校验 API Key 非空，占位符会反复重启
+# 注意：必须同时被 deriver/dreamer/embedding 复用，否则 specialist 会回落 openai.com
 # ============================================================
-HONCHO_LLM_BASE_URL=https://api.openai.com/v1
+HONCHO_LLM_BASE_URL=https://api.siliconflow.cn/v1
 
 # ============================================================
 # 敏感值（必填，env.cfg 不提交）— 在 .gitignore 已忽略
diff --git a/honcho/stack.yml b/honcho/stack.yml
index bada9fc..486ffb4 100644
--- a/honcho/stack.yml
+++ b/honcho/stack.yml
@@ -93,7 +93,14 @@ services:
       - CACHE_ENABLED=true
       - NAMESPACE=${HONCHO_NAMESPACE:-honcho}
       - DERIVER_MODEL_CONFIG__TRANSPORT=openai
-      - DERIVER_MODEL_CONFIG__MODEL=${HONCHO_DERIVER_MODEL:-gpt-5.4-mini}
+      - DERIVER_MODEL_CONFIG__MODEL=${HONCHO_DERIVER_MODEL:-deepseek-ai/DeepSeek-V4-Flash}
+      - DERIVER_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1}
+      - DEDUCTION_MODEL_CONFIG__TRANSPORT=openai
+      - DEDUCTION_MODEL_CONFIG__MODEL=${HONCHO_DEDUCTION_MODEL:-deepseek-ai/DeepSeek-V4-Flash}
+      - DEDUCTION_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1}
+      - INDUCTION_MODEL_CONFIG__TRANSPORT=openai
+      - INDUCTION_MODEL_CONFIG__MODEL=${HONCHO_INDUCTION_MODEL:-deepseek-ai/DeepSeek-V4-Flash}
+      - INDUCTION_MODEL_CONFIG__OVERRIDES__BASE_URL=${HONCHO_LLM_BASE_URL:-https://api.siliconflow.cn/v1}
       - DERIVER_WORKERS=${HONCHO_DERIVER_WORKERS:-1}
       - EMBEDDING_MODEL_CONFIG__TRANSPORT=openai
       - EMBEDDING_MODEL_CONFIG__MODEL=${HONCHO_EMBEDDING_MODEL:-Qwen/Qwen3-Embedding-8B}