diff --git a/docker-compose.mtp.yaml b/docker-compose.mtp.yaml
index a74f116857..880307c089 100644
--- a/docker-compose.mtp.yaml
+++ b/docker-compose.mtp.yaml
@@ -25,8 +25,21 @@ services:
       # was raised to match the README since Gemma-4's hybrid SWA
       # architecture (most layers use a small fixed attention window, only
       # every 5th layer is full-context) keeps KV cache growth with n_ctx
-      # much cheaper than a plain transformer's.
-      - 'MTP_CTX_SIZE=32768'
+      # much cheaper than a plain transformer's -- measured at only 271 MiB
+      # of KV cache (target + draft) per 32768-token slot.
+      #
+      # -np 4 gives open-webui's browser chat and the voice assistant (and
+      # anything else hitting this OpenAI-compatible endpoint) their own
+      # slot each instead of one contending for a single shared slot --
+      # with -np 1, whichever of them touched the slot last would evict the
+      # other's live KV cache, forcing a full prompt reprocess (multi-second
+      # stall) on the next request from the loser. n_ctx_slot = n_ctx /
+      # n_parallel, so total ctx is quadrupled to keep each slot at the same
+      # 32768 budget as before -- 131072 matches this model's native
+      # n_ctx_train exactly, and only costs ~810 MiB more KV cache VRAM
+      # than the old -np 1 config (still well under the ~7GB that's free).
+      - 'MTP_CTX_SIZE=131072'
+      - 'MTP_N_PARALLEL=4'
       - 'MTP_BATCH_SIZE=512'
       # Server-side default -- only applies when the client doesn't send its
       # own temperature (Open WebUI won't unless you set it in the model's