diff --git a/examples/llava_onevision/conf/serve/7b.yaml b/examples/llava_onevision/conf/serve/7b.yaml
index 65112005a2..fbfc1557a2 100644
--- a/examples/llava_onevision/conf/serve/7b.yaml
+++ b/examples/llava_onevision/conf/serve/7b.yaml
@@ -6,7 +6,7 @@
     pipeline_parallel_size: 1
     gpu_memory_utilization: 0.9
     max_model_len: 32768
-    limit_mm_per_prompt: image=8
+    limit_mm_per_prompt: '{"image": 8}'
     max_num_seqs: 16
     enforce_eager: true
     trust_remote_code: true
diff --git a/examples/minicpm_o_2.6/conf/serve/7b.yaml b/examples/minicpm_o_2.6/conf/serve/7b.yaml
index 6d5fa06acc..7edf0f8e77 100644
--- a/examples/minicpm_o_2.6/conf/serve/7b.yaml
+++ b/examples/minicpm_o_2.6/conf/serve/7b.yaml
@@ -5,7 +5,7 @@
     pipeline_parallel_size: 1
     gpu_memory_utilization: 0.9
     max_num_seqs: 256
-    limit_mm_per_prompt: image=18
+    limit_mm_per_prompt: '{"image": 18}'
     port: 9010
     trust_remote_code: true
     enable_chunked_prefill: true
diff --git a/examples/robobrain/conf/serve/7b.yaml b/examples/robobrain/conf/serve/7b.yaml
index 8505a31cff..68592882d0 100644
--- a/examples/robobrain/conf/serve/7b.yaml
+++ b/examples/robobrain/conf/serve/7b.yaml
@@ -6,7 +6,7 @@
     pipeline_parallel_size: 1
     gpu_memory_utilization: 0.9
     max_model_len: 32768
-    limit_mm_per_prompt: image=8
+    limit_mm_per_prompt: '{"image": 8}'
     max_num_seqs: 16
     trust_remote_code: true
     enable_chunked_prefill: false
diff --git a/examples/robobrain2/conf/serve/32b.yaml b/examples/robobrain2/conf/serve/32b.yaml
index 22d461d932..5a0c0fe000 100644
--- a/examples/robobrain2/conf/serve/32b.yaml
+++ b/examples/robobrain2/conf/serve/32b.yaml
@@ -7,7 +7,7 @@
     pipeline_parallel_size: 1
     max_num_seqs: 8 # Even at full 32,768 context usage, 8 concurrent operations won't trigger OOM
     gpu_memory_utilization: 0.9
-    limit_mm_per_prompt: image=18 # should be customized, 18 images/request is enough for most scenarios
+    limit_mm_per_prompt: '{"image": 18}' # should be customized, 18 images/request is enough for most scenarios
     port: 9010
     trust_remote_code: true
     enforce_eager: false # set true if use FlagGems
diff --git a/examples/robobrain2/conf/serve/3b.yaml b/examples/robobrain2/conf/serve/3b.yaml
index d378a87c8e..1d7cdf3088 100644
--- a/examples/robobrain2/conf/serve/3b.yaml
+++ b/examples/robobrain2/conf/serve/3b.yaml
@@ -7,7 +7,7 @@
     pipeline_parallel_size: 1
     max_num_seqs: 8 # Even at full 32,768 context usage, 8 concurrent operations won't trigger OOM
     gpu_memory_utilization: 0.9
-    limit_mm_per_prompt: image=18 # should be customized, 18 images/request is enough for most scenarios
+    limit_mm_per_prompt: '{"image": 18}' # should be customized, 18 images/request is enough for most scenarios
     port: 9010
     trust_remote_code: true
     enforce_eager: false # set true if use FlagGems
diff --git a/examples/robobrain2/conf/serve/7b.yaml b/examples/robobrain2/conf/serve/7b.yaml
index ff96872566..f62ecea9dc 100644
--- a/examples/robobrain2/conf/serve/7b.yaml
+++ b/examples/robobrain2/conf/serve/7b.yaml
@@ -7,7 +7,7 @@
     pipeline_parallel_size: 1
     max_num_seqs: 8 # Even at full 32,768 context usage, 8 concurrent operations won't trigger OOM
     gpu_memory_utilization: 0.9
-    limit_mm_per_prompt: image=18 # should be customized, 18 images/request is enough for most scenarios
+    limit_mm_per_prompt: '{"image": 18}' # should be customized, 18 images/request is enough for most scenarios
     port: 9010
     trust_remote_code: true
     enforce_eager: false # set true if use FlagGems