diff --git a/hosts/focalor/modules/strata.nix b/hosts/focalor/modules/strata.nix
index 6fc30ed..f0e51f3 100644
--- a/hosts/focalor/modules/strata.nix
+++ b/hosts/focalor/modules/strata.nix
@@ -16,7 +16,10 @@ let
   cfg = config.modules.strata;
   port = 8092;
   data = "/data/strata-data";
-  model = "${data}/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS";
+  # niri's voice: a LoRA trained on her DMs with ana (~290 dense attention/GDN/shared-expert tensors), merged by
+  # requantizing those tensors from the bf16 weights plus the delta to their own GSQ-RCO types; experts unchanged.
+  # Held-out DM NLL 2.544 -> 2.341 in strata. The unmodified model is in models/IQ3_XXS.
+  model = "${data}/models/IQ3_XXS-bf16-niri-e1/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS";
   cuda = pkgs.cudaPackages_13_0; # Strata pins CUDA 13.0: 13.2.0/13.2.1 miscompile for sm_120
   llamaCpp = pkgs.fetchFromGitHub {
     # the ggml and mtmd sources setup.py pins (LLAMA_CPP_COMMIT)