Revert active model to Qwen3-8B (MiniCPM incompatible with this env)

MiniCPM4.1-8B's bundled code targets transformers 4.56, but Gradio 6.18 forces
huggingface-hub>=1.0 and thus transformers 5.x, which refactored the internals
that 4.56-era remote code relies on (the is_torch_fx_available shim only covers
the first of many breakages). Qwen3-8B is natively supported in transformers 5.x
with no remote code, so it loads cleanly in this Space. Shim kept harmlessly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app.py

@@ -43,13 +43,17 @@ from shards import SHARD_ORDER, SHARDS
 
 # ---------------------------------------------------------------------------- #
 # Model — keep the id in one easy-to-swap constant.
-#   Active:   openbmb/MiniCPM4.1-8B  (<=32B; qualifies for the OpenBMB sponsor prize)
-#   Swap-in:  Qwen/Qwen3-8B          (strong dialogue, ZeroGPU-friendly, LoRA-ready)
+#   Active:   Qwen/Qwen3-8B          (natively supported by transformers 5.x; no
+#                                     remote code; works with this Gradio 6.18 env)
 #   Step up:  Qwen/Qwen3-14B         (if replies feel thin and compute allows)
-# Both MiniCPM4.1 and Qwen3 are hybrid reasoning models that honor
-# apply_chat_template(enable_thinking=False) for snappy, non-thinking replies.
+# Note: openbmb/MiniCPM4.1-8B (the OpenBMB sponsor-prize model) is NOT usable
+# here — its bundled modeling code targets transformers 4.56, but Gradio 6.18
+# forces huggingface-hub>=1.0 and therefore transformers 5.x, which refactored
+# the internals that 4.56-era remote code depends on. Running it would require
+# downgrading Gradio. The is_torch_fx_available shim above is kept harmlessly in
+# case a future, 5.x-compatible MiniCPM revision is published.
 # ---------------------------------------------------------------------------- #
-MODEL_ID = "openbmb/MiniCPM4.1-8B"
+MODEL_ID = "Qwen/Qwen3-8B"
 
 MODEL = None
 TOKENIZER = None