Batch model test w streaming, thinking + synthesis

Interactive Ollama batch tester that streams responses—including "thinking" blocks when supported—across either a custom model list or all locally available models. Features live output per model, automatic fallback if target models missing, and optional multi-response synthesis (commented out). Runs in a REPL loop until user exits with "X". Designed for rapid comparative evaluation of local LLMs.

examples/streaming_and_thinking_batch_model_output_comparison.py

@@ -0,0 +1,107 @@
+from ollama import Client
+
+
+def batch_test_ollama_models():
+    client = Client()
+    
+    try:
+        available_models_info = client.list()
+        available_models = [m.model if hasattr(m, 'model') else m.get('model', m.get('name', '')) for m in available_models_info.get('models', [])]
+    except Exception as e:
+        print(f"Error listing models: {e}")
+        available_models = []
+    
+    # Custom list of desired models to test - filtered by availability
+    models_to_compare = [
+        "gemini-3-flash-preview:latest",
+        "llama3.1:latest",
+        "gemma:latest",
+        "qwen3:latest",
+        "devstral-2:123b-cloud",
+    ]
+    
+    # Verify models are actually installed/registered
+    final_model_list = [m for m in models_to_compare if m in available_models]
+
+    # Or use all available models excluding a list of models to exclude
+    # models = [m for m in available_models if m not in ["model1", "model2", "model3"]]
+
+    if not final_model_list and available_models:
+        print("None of the target models are installed. Using all available model for testing.")
+        final_model_list = available_models
+    elif not available_models:
+        print("No models found in Ollama. Please pull some models first.")
+        exit(1)
+    
+    synthesis_model = final_model_list[0] # Set to a long-model else default to first in list
+    
+    responses = []  # List of strings for each model
+    synthesized_history = [] # History of final responses
+    
+    while True:
+        prompt = input("\nEnter prompt or X to quit: ")
+        if prompt.upper() == "X":
+            break
+           
+        # Reset responses for new prompt cycle
+        responses = [[] for _ in range(len(final_model_list))]
+
+        for i, model in enumerate(final_model_list):
+            print(f"\n--- Model [{i+1}/{len(final_model_list)}]: {model} ---")
+            in_thinking = False
+            try:
+                for part in client.chat(model=model, messages=[{"role": "user", "content": prompt}], stream=True):
+                    
+                    # Handle thinking block if present
+                    if hasattr(part.message, 'thinking') and part.message.thinking:
+                        if not in_thinking:
+                            print("\nThinking: ", end="", flush=True)
+                            in_thinking = True
+                        print(part.message.thinking, end="", flush=True)
+                        # [i].append(part.message.thinking) # Don't append thinking to responses
+
+                    # Handle actual content
+                    if part.message.content:
+                        if in_thinking:
+                            print("\nContent: ", end="", flush=True)
+                            in_thinking = False
+                        print(part.message.content, end="", flush=True)
+                        responses[i].append(part.message.content) # Collect responses to save to file or do synthesis/summary/evaluation on all of them
+            except Exception as e:
+                print(f"\nError with model {model}: {e}")
+        
+        print("\n\n" + "="*50)
+        print("SYNTHESIZING RESPONSES using", synthesis_model)
+        print("="*50 + "\n")
+
+        # Prepare data for synthesis
+        full_responses_text = ""
+
+        for idx, model_name in enumerate(final_model_list):
+            resp_text = "".join(responses[idx])
+            full_responses_text += f"""
+            Model: {model_name}
+            Response: {resp_text}
+            """
+
+        # PROMPT FOR PROCESSING ALL RESPONSES TOGETHER - Edit as you see fit
+        synthesis_prompt = f"Here are several outputs for the following input: '{prompt}':\n\n{full_responses_text}\nSynthesize these outputs into one distilled but comprehensive, high-quality response that consolidates all of the unique points in their best form from all of the outputs, avoiding redundancy and optimizing for quality, clarity and effectiveness: "
+
+        current_synthesis = []
+        try:
+            for part in client.chat(model=synthesis_model, messages=[{"role": "user", "content": synthesis_prompt}], stream=True):
+                if hasattr(part.message, 'thinking') and part.message.thinking:
+                    print(part.message.thinking, end="", flush=True)
+                if part.message.content:
+                    print(part.message.content, end="", flush=True)
+                    current_synthesis.append(part.message.content)
+            synthesized_history.append("".join(current_synthesis))
+        except Exception as e:
+            print(f"\nError during synthesis: {e}")
+        print("\n" + "-"*50 + "\n")
+
+def main():
+    batch_test_ollama_models()
+
+if __name__ == "__main__":
+    main()