Update app_deepseek.py

app_deepseek.py

@@ -389,7 +389,7 @@ SEARCH_LANGUAGE_CYCLE = ["arabic", "english"]
 # "hf": the old transformers path (only kept for reference; on TPU it produced ~3 chars/s and hung).
 BACKEND = os.environ.get("GIJL_BACKEND", "llamacpp").strip().lower()
 GGUF_REPO = os.environ.get("GIJL_GGUF_REPO", "unsloth/DeepSeek-V4-Flash-0731-GGUF")
-GGUF_QUANT = os.environ.get("GIJL_GGUF_QUANT", "Q8_0")   # highest quant unsloth ships; choose_quant() steps this down
+GGUF_QUANT = os.environ.get("GIJL_GGUF_QUANT", "UD-IQ3_XXS")   # highest quant unsloth ships; choose_quant() steps this down
                                                            # automatically to whatever fits the detected VRAM/RAM
                                                            # budget (see _memory_budget_gb) -- set GIJL_GGUF_QUANT
                                                            # explicitly to pin a smaller quant instead.