diff --git a/recipes/glm-5-3-flash-low-effort-on-2x-cmp170hx-exl3-tp2-four-slots/recipe.yaml b/recipes/glm-5-3-flash-low-effort-on-2x-cmp170hx-exl3-tp2-four-slots/recipe.yaml
index 8a76d6bb7..1c7b1a323 100644
--- a/recipes/glm-5-3-flash-low-effort-on-2x-cmp170hx-exl3-tp2-four-slots/recipe.yaml
+++ b/recipes/glm-5-3-flash-low-effort-on-2x-cmp170hx-exl3-tp2-four-slots/recipe.yaml
@@ -45,10 +45,13 @@ config:
     GLM_PERSISTENT_CACHE_DIR: /absolute/path/to/session-cache
     GLM_PERSISTENT_CACHE_GIB: '50'
     GLM_PERSISTENT_CACHE_MIN_FREE_GIB: '16'
+    RECURRENT_STATE_DTYPE: fp32
   client_model: GLM-5.3-Flash-low
   native_layout: '4 active sequences, row32 verifier, K7 DFlash2 plus hybrid15, native broadcast/gather with NCCL reductions'
   request_defaults:
     reasoning_effort: low
+  runtime:
+    notes: 'RECURRENT_STATE_DTYPE sets the recurrent-state storage dtype: fp32 default (launch adds --recurrent_state_fp32); fp16 opts into the upstream FP16 recurrent storage patch (0005-upstream-fp16-recurrent-state); accepts only fp16|fp32, anything else exits 2 at launch.'
 performance:
   claimed_by: submitter
   tok_s: 41.721784744869