Update app.py

app.py

@@ -21,14 +21,49 @@ from transformers import (
 
 # Small / moderate models that work with AutoModelForCausalLM
 MODEL_CHOICES = [
-    "distilgpt2",                       # tiny baseline
-    "sshleifer/tiny-gpt2",              # toy
-    "Qwen/Qwen2.5-0.5B-Instruct",       # nice small instruct model (GPU better, but can try CPU)
+    # Very small / light (good for CPU Spaces)
+    "distilgpt2",
+    "gpt2",
+    "sshleifer/tiny-gpt2",
+    "LiquidAI/LFM2-350M",
+    "google/gemma-3-270m-it",
+    "Qwen/Qwen2.5-0.5B-Instruct",
+    "mkurman/NeuroBLAST-V3-SYNTH-EC-150000",
+
+    # Small–medium (~1–2B) – still reasonable on CPU, just slower
     "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
     "google/gemma-3-1b-it",
+    "meta-llama/Llama-3.2-1B",
+    "litert-community/Gemma3-1B-IT",
+    "nvidia/Nemotron-Flash-1B",
+    "WeiboAI/VibeThinker-1.5B",
+    "Qwen/Qwen3-1.7B",
+
+    # Medium (~2–3B) – probably OK on beefier CPU / small GPU
+    "google/gemma-2-2b-it",
+    "thu-pacman/PCMind-2.1-Kaiyuan-2B",
+    "opendatalab/MinerU-HTML",            # 0.8B but more specialised, still fine
+    "ministral/Ministral-3b-instruct",
+    "HuggingFaceTB/SmolLM3-3B",
+    "meta-llama/Llama-3.2-3B-Instruct",
+    "nvidia/Nemotron-Flash-3B-Instruct",
+    "Qwen/Qwen2.5-3B-Instruct",
+
+    # Heavier (4–8B) – you really want a GPU Space for these
+    "Qwen/Qwen3-4B",
+    "Qwen/Qwen3-4B-Thinking-2507",
+    "Qwen/Qwen3-4B-Instruct-2507",
+    "mistralai/Mistral-7B-Instruct-v0.2",
+    "allenai/Olmo-3-7B-Instruct",
+    "Qwen/Qwen2.5-7B-Instruct",
+    "meta-llama/Meta-Llama-3-8B-Instruct",
+    "meta-llama/Llama-3.1-8B",
+    "meta-llama/Llama-3.1-8B-Instruct",
+    "openbmb/MiniCPM4.1-8B",
+    "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
+    "rl-research/DR-Tulu-8B",
 ]
-
-DEFAULT_MODEL = "distilgpt2"  # safe default for CPU Space
+DEFAULT_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"  # or TinyLlama, or stick with distilgpt2
 
 device = 0 if torch.cuda.is_available() else -1