diff --git a/include/agent_config.h b/include/agent_config.h
index 32b28aa..15d8774 100644
--- a/include/agent_config.h
+++ b/include/agent_config.h
@@ -249,6 +249,9 @@
 #define AGENT_CFG_KEY_FEISHU_USER_TOKEN "feishu_user_token"
 #define AGENT_CFG_KEY_LLM_HOST "llm_host"
 #define AGENT_CFG_KEY_LLM_PATH "llm_path"
+#define AGENT_CFG_KEY_VISION_MODEL "vision_model"
+#define AGENT_CFG_KEY_VISION_HOST "vision_host"
+#define AGENT_CFG_KEY_VISION_API_KEY "vision_api_key"
 #define AGENT_CFG_KEY_GATEWAY_HOST "gateway_host"
 #define AGENT_CFG_KEY_GATEWAY_PORT "gateway_port"
 #define AGENT_CFG_KEY_GATEWAY_TOKEN "gateway_token"
diff --git a/src/channels/cmd_llm.c b/src/channels/cmd_llm.c
index a77cbf7..e7c2641 100644
--- a/src/channels/cmd_llm.c
+++ b/src/channels/cmd_llm.c
@@ -47,7 +47,7 @@ static const router_preset_t g_router_presets[] = {
     { "glm", "open.bigmodel.cn", "/api/paas/v4/chat/completions", "glm-4-flash", 1 },
     { "openai", "api.openai.com", "/v1/chat/completions", "gpt-4o", 3 },
     { "claude", "api.anthropic.com", "/v1/messages", "claude-sonnet-4-20250514", 3 },
-    { "mimo", "api.xiaomimimo.com", "/v1/chat/completions", "MiMo-v2-Flash", 1 },
+    { "mimo", "api.xiaomimimo.com", "/v1/chat/completions", "mimo-v2-flash", 1 },
     { "openrouter", "openrouter.ai", "/api/v1/chat/completions",
         "openrouter/hunter-alpha", 2 },
     { NULL, NULL, NULL, NULL, 0 }
@@ -223,6 +223,85 @@ void cmd_set_llm(int argc, char** argv)
 }
 
 
+/* ── cmd_set_vision_llm ──────────────────────────────────────── */
+
+void cmd_set_vision_llm(int argc, char** argv)
+{
+    if (argc < 2) {
+        printf("Usage: set_vision_llm <preset> [api_key]\n"
+               "       set_vision_llm <host> <model> [api_key]\n"
+               "       set_vision_llm clear\n"
+               "\n"
+               "Set an independent vision model. If not configured,\n"
+               "vision calls inherit the main LLM config.\n"
+               "\n"
+               "Presets:\n"
+               "  mimo     - api.xiaomimimo.com  (mimo-v2-omni)\n"
+               "  openai   - api.openai.com  (gpt-4o)\n"
+               "  qwen     - dashscope.aliyuncs.com  (qwen-vl-max)\n"
+               "  glm      - open.bigmodel.cn  (glm-4v-flash)\n"
+               "\n"
+               "Examples:\n"
+               "  set_vision_llm mimo <api_key>\n"
+               "  set_vision_llm clear\n");
+        return;
+    }
+
+    const char* arg1 = argv[1];
+
+    /* clear: remove vision-specific config, fall back to main LLM */
+    if (strcmp(arg1, "clear") == 0) {
+        llm_set_vision_model(NULL, NULL, NULL);
+        printf("Vision LLM config cleared (using main LLM).\n");
+        return;
+    }
+
+    const char* host = NULL;
+    const char* model = NULL;
+    const char* api_key = NULL;
+
+    /* Check if arg1 is a preset name */
+    static const struct {
+        const char* name;
+        const char* host;
+        const char* model;
+    } vision_presets[] = {
+        { "mimo", "api.xiaomimimo.com", "mimo-v2-omni" },
+        { "openai", "api.openai.com", "gpt-4o" },
+        { "qwen", "dashscope.aliyuncs.com", "qwen-vl-max" },
+        { "glm", "open.bigmodel.cn", "glm-4v-flash" },
+        { NULL, NULL, NULL }
+    };
+
+    bool found = false;
+    for (int i = 0; vision_presets[i].name; i++) {
+        if (strcmp(vision_presets[i].name, arg1) == 0) {
+            host = vision_presets[i].host;
+            model = vision_presets[i].model;
+            if (argc >= 3)
+                api_key = argv[2];
+            found = true;
+            break;
+        }
+    }
+
+    if (!found) {
+        /* Custom: arg1 = host, arg2 = model, arg3 = api_key */
+        host = arg1;
+        if (argc >= 3)
+            model = argv[2];
+        if (argc >= 4)
+            api_key = argv[3];
+    }
+
+    llm_set_vision_model(host, model, api_key);
+
+    printf("Vision LLM: %s (model: %s)\n", host, model ? model : "(inherit)");
+    if (api_key)
+        printf("Vision API key saved.\n");
+}
+
+
 /* ── list_models helpers ──────────────────────────────────────── */
 
 #define LIST_MODELS_BUF_SIZE (256 * 1024)
diff --git a/src/channels/cmd_llm.h b/src/channels/cmd_llm.h
index 8fb5232..a9a0378 100644
--- a/src/channels/cmd_llm.h
+++ b/src/channels/cmd_llm.h
@@ -17,6 +17,7 @@
 #pragma once
 
 void cmd_set_llm(int argc, char** argv);
+void cmd_set_vision_llm(int argc, char** argv);
 void cmd_list_models(int argc, char** argv);
 void cmd_router_status(void);
 void cmd_router_profile(int argc, char** argv);
diff --git a/src/channels/nsh_commands.c b/src/channels/nsh_commands.c
index f750c4e..defb312 100644
--- a/src/channels/nsh_commands.c
+++ b/src/channels/nsh_commands.c
@@ -96,6 +96,7 @@ static void cmd_help(void)
         "  set_feishu_app <app_id> <app_secret>  - Set Feishu app credentials\n"
         "  set_feishu_user_token <token>  - Set Feishu user_access_token for doc APIs\n"
         "  set_llm <preset|host> [model] [key] - Switch LLM backend (kimi/qwen/deepseek/glm/openai)\n"
+        "  set_vision_llm <preset|host> [model] [key] - Set independent vision model\n"
         "  list_models [--free] [keyword] - List available models (openrouter)\n"
         "  memory_read          - Read MEMORY.md\n"
         "  memory_write <text>  - Write MEMORY.md (quote text)\n"
@@ -462,6 +463,9 @@ static void cmd_config_show(void)
     SHOW_CFG("Model", AGENT_CFG_KEY_MODEL, false);
     SHOW_CFG("LLM Host", AGENT_CFG_KEY_LLM_HOST, false);
     SHOW_CFG("LLM Path", AGENT_CFG_KEY_LLM_PATH, false);
+    SHOW_CFG("Vision Model", AGENT_CFG_KEY_VISION_MODEL, false);
+    SHOW_CFG("Vision Host", AGENT_CFG_KEY_VISION_HOST, false);
+    SHOW_CFG("Vision Key", AGENT_CFG_KEY_VISION_API_KEY, true);
     SHOW_CFG("Proxy Host", AGENT_CFG_KEY_PROXY_HOST, false);
     SHOW_CFG("Proxy Port", AGENT_CFG_KEY_PROXY_PORT, false);
     SHOW_CFG("SerpAPI Key", AGENT_CFG_KEY_SERP_KEY, true);
@@ -876,6 +880,8 @@ static void* cli_thread(void* arg)
             cmd_set_feishu_user_token(argc, argv);
         else if (strcmp(cmd, "set_llm") == 0)
             cmd_set_llm(argc, argv);
+        else if (strcmp(cmd, "set_vision_llm") == 0)
+            cmd_set_vision_llm(argc, argv);
         else if (strcmp(cmd, "list_models") == 0)
             cmd_list_models(argc, argv);
         else if (strcmp(cmd, "memory_read") == 0)
diff --git a/src/llm/llm_proxy.c b/src/llm/llm_proxy.c
index c682d90..48157cc 100644
--- a/src/llm/llm_proxy.c
+++ b/src/llm/llm_proxy.c
@@ -46,6 +46,11 @@ static char s_llm_host[128] = AGENT_LLM_API_HOST;
 static char s_llm_path[128] = AGENT_LLM_API_PATH;
 static char s_llm_port[8] = "443"; /* "443" for HTTPS, "80" etc for HTTP */
 
+/* Independent vision model config — empty means fall back to main LLM */
+static char s_vision_model[64] = { 0 };
+static char s_vision_host[128] = { 0 };
+static char s_vision_api_key[128] = { 0 };
+
 static pthread_mutex_t s_llm_lock = PTHREAD_MUTEX_INITIALIZER;
 
 /* Check if host uses OpenAI-compatible max_completion_tokens param */
@@ -124,6 +129,17 @@ int llm_proxy_init(void)
     if (claw_config_get("llm_port", tmp, sizeof(tmp)) == OK && tmp[0])
         strncpy(s_llm_port, tmp, sizeof(s_llm_port) - 1);
 
+    /* Load optional vision model config — empty falls back to main LLM */
+    memset(tmp, 0, sizeof(tmp));
+    if (claw_config_get(AGENT_CFG_KEY_VISION_MODEL, tmp, sizeof(tmp)) == OK && tmp[0])
+        strncpy(s_vision_model, tmp, sizeof(s_vision_model) - 1);
+    memset(tmp, 0, sizeof(tmp));
+    if (claw_config_get(AGENT_CFG_KEY_VISION_HOST, tmp, sizeof(tmp)) == OK && tmp[0])
+        strncpy(s_vision_host, tmp, sizeof(s_vision_host) - 1);
+    memset(tmp, 0, sizeof(tmp));
+    if (claw_config_get(AGENT_CFG_KEY_VISION_API_KEY, tmp, sizeof(tmp)) == OK && tmp[0])
+        strncpy(s_vision_api_key, tmp, sizeof(s_vision_api_key) - 1);
+
     if (s_api_key[0])
         syslog(LOG_INFO, "[%s] LLM proxy initialized (model: %s, host: %s)\n", TAG,
             s_model, s_llm_host);
@@ -212,6 +228,78 @@ void llm_snapshot_config(char* model, size_t model_sz,
     pthread_mutex_unlock(&s_llm_lock);
 }
 
+/* ── Vision config snapshot ──────────────────────────────────── */
+
+void llm_snapshot_vision_config(char* model, size_t model_sz,
+    char* api_key, size_t key_sz,
+    char* host, size_t host_sz)
+{
+    pthread_mutex_lock(&s_llm_lock);
+
+    /* Use vision-specific config when set, otherwise fall back to main LLM */
+    if (s_vision_model[0])
+        strncpy(model, s_vision_model, model_sz - 1);
+    else
+        strncpy(model, s_model, model_sz - 1);
+    model[model_sz - 1] = '\0';
+
+    if (s_vision_api_key[0])
+        strncpy(api_key, s_vision_api_key, key_sz - 1);
+    else
+        strncpy(api_key, s_api_key, key_sz - 1);
+    api_key[key_sz - 1] = '\0';
+
+    if (s_vision_host[0])
+        strncpy(host, s_vision_host, host_sz - 1);
+    else
+        strncpy(host, s_llm_host, host_sz - 1);
+    host[host_sz - 1] = '\0';
+
+    pthread_mutex_unlock(&s_llm_lock);
+}
+
+/* ── Vision model setter ─────────────────────────────────────── */
+
+int llm_set_vision_model(const char* host, const char* model,
+    const char* api_key)
+{
+    pthread_mutex_lock(&s_llm_lock);
+
+    if (host && host[0]) {
+        claw_config_set(AGENT_CFG_KEY_VISION_HOST, host);
+        strncpy(s_vision_host, host, sizeof(s_vision_host) - 1);
+        s_vision_host[sizeof(s_vision_host) - 1] = '\0';
+    } else {
+        claw_config_set(AGENT_CFG_KEY_VISION_HOST, "");
+        s_vision_host[0] = '\0';
+    }
+
+    if (model && model[0]) {
+        claw_config_set(AGENT_CFG_KEY_VISION_MODEL, model);
+        strncpy(s_vision_model, model, sizeof(s_vision_model) - 1);
+        s_vision_model[sizeof(s_vision_model) - 1] = '\0';
+    } else {
+        claw_config_set(AGENT_CFG_KEY_VISION_MODEL, "");
+        s_vision_model[0] = '\0';
+    }
+
+    if (api_key && api_key[0]) {
+        claw_config_set(AGENT_CFG_KEY_VISION_API_KEY, api_key);
+        strncpy(s_vision_api_key, api_key, sizeof(s_vision_api_key) - 1);
+        s_vision_api_key[sizeof(s_vision_api_key) - 1] = '\0';
+    } else {
+        claw_config_set(AGENT_CFG_KEY_VISION_API_KEY, "");
+        s_vision_api_key[0] = '\0';
+    }
+
+    syslog(LOG_INFO, "[%s] Vision LLM config updated: model=%s host=%s\n",
+        TAG, s_vision_model[0] ? s_vision_model : "(inherit)",
+        s_vision_host[0] ? s_vision_host : "(inherit)");
+
+    pthread_mutex_unlock(&s_llm_lock);
+    return OK;
+}
+
 /* ── HTTP helpers ─────────────────────────────────────────── */
 
 /**
diff --git a/src/llm/llm_proxy.h b/src/llm/llm_proxy.h
index dde3169..b92d012 100644
--- a/src/llm/llm_proxy.h
+++ b/src/llm/llm_proxy.h
@@ -94,6 +94,18 @@ int llm_chat_vision_raw(const char* prompt,
     const char* mime_type,
     char* response_buf, size_t buf_size);
 
+/** Vision config snapshot: returns the vision-specific model/api_key/host.
+ *  If vision model is not configured, falls back to the main LLM config. */
+void llm_snapshot_vision_config(char* model, size_t model_sz,
+    char* api_key, size_t key_sz,
+    char* host, size_t host_sz);
+
+/** Set independent vision model config.
+ *  host/model/api_key: set non-NULL/non-empty to override, NULL to clear.
+ *  Empty vision config falls back to main LLM config automatically. */
+int llm_set_vision_model(const char* host, const char* model,
+    const char* api_key);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/llm/llm_vision.c b/src/llm/llm_vision.c
index fa337a5..549a15b 100644
--- a/src/llm/llm_vision.c
+++ b/src/llm/llm_vision.c
@@ -43,7 +43,7 @@ int llm_chat_vision(const char* prompt, const char* image_b64,
     size_t buf_size)
 {
     char model[64], api_key[128], llm_host[128];
-    llm_snapshot_config(model, sizeof(model),
+    llm_snapshot_vision_config(model, sizeof(model),
         api_key, sizeof(api_key),
         llm_host, sizeof(llm_host));
 
@@ -215,7 +215,7 @@ int llm_chat_vision_raw(const char* prompt,
     char* response_buf, size_t buf_size)
 {
     char model[64], api_key[128], llm_host[128];
-    llm_snapshot_config(model, sizeof(model),
+    llm_snapshot_vision_config(model, sizeof(model),
         api_key, sizeof(api_key),
         llm_host, sizeof(llm_host));