Add Qwen3.8 Max to live bench

lcb_runner/lm_styles.py

@@ -12,6 +12,7 @@ class LMStyle(Enum):
     CohereCommand = "CohereCommand"
     DataBricks = "DataBricks"
     DeepSeekAPI = "DeepSeekAPI"
+    QwenAPI = "QwenAPI"
 
     GenericBase = "GenericBase"
 
@@ -544,6 +545,13 @@ def __hash__(self) -> int:
         datetime(2024, 3, 31),
         link="https://huggingface.co/qwen/Qwen1.5-72B-Chat/",
     ),
+    LanguageModel(
+        "qwen3.8-max",
+        "Qwen3.8-Max",
+        LMStyle.QwenAPI,
+        datetime(2026, 8, 2),
+        link="https://qwen.ai/blog?id=qwen3.8",
+    ),
     LanguageModel(
         "abacusai/Smaug-2-72B",
         "Smaug-2-72B ",

lcb_runner/prompts/code_generation.py

@@ -206,7 +206,7 @@ def get_example_prompt(example):
 def format_prompt_generation(
     question: CodeGenerationProblem, LanguageModelStyle: LMStyle
 ) -> str:
-    if LanguageModelStyle in [LMStyle.OpenAIChat, LMStyle.DeepSeekAPI]:
+    if LanguageModelStyle in [LMStyle.OpenAIChat, LMStyle.DeepSeekAPI, LMStyle.QwenAPI]:
         chat_messages = [
             {
                 "role": "system",

lcb_runner/runner/qwen_runner.py

@@ -0,0 +1,70 @@
+import os
+from time import sleep
+
+try:
+    import openai
+    from openai import OpenAI
+except ImportError as e:
+    pass
+
+from lcb_runner.runner.base_runner import BaseRunner
+
+
+class QwenRunner(BaseRunner):
+    client = OpenAI(
+        api_key=os.getenv("DASHSCOPE_API"),
+        base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+    )
+
+    def __init__(self, args, model):
+        super().__init__(args, model)
+        self.client_kwargs: dict[str | str] = {
+            "model": args.model,
+            "temperature": args.temperature,
+            "max_tokens": args.max_tokens,
+            "top_p": args.top_p,
+            "frequency_penalty": 0,
+            "presence_penalty": 0,
+            "n": 1,
+            "timeout": args.openai_timeout,
+            # "stop": args.stop, --> stop is only used for base models currently
+        }
+
+    def _run_single(self, prompt: list[dict[str, str]]) -> list[str]:
+        assert isinstance(prompt, list)
+
+        def __run_single(counter):
+            try:
+                response = self.client.chat.completions.create(
+                    messages=prompt,
+                    **self.client_kwargs,
+                )
+                content = response.choices[0].message.content
+                return content
+            except (
+                openai.APIError,
+                openai.RateLimitError,
+                openai.InternalServerError,
+                openai.OpenAIError,
+                openai.APIStatusError,
+                openai.APITimeoutError,
+                openai.InternalServerError,
+                openai.APIConnectionError,
+            ) as e:
+                print("Exception: ", repr(e))
+                print("Sleeping for 30 seconds...")
+                print("Consider reducing the number of parallel processes.")
+                sleep(30)
+                return QwenRunner._run_single(prompt)
+            except Exception as e:
+                print(f"Failed to run the model for {prompt}!")
+                print("Exception: ", repr(e))
+                raise e
+
+        outputs = []
+        try:
+            for _ in range(self.args.n):
+                outputs.append(__run_single(10))
+        except Exception as e:
+            raise e
+        return outputs

lcb_runner/runner/runner_utils.py

@@ -30,6 +30,10 @@ def build_runner(args, model: LanguageModel):
         from lcb_runner.runner.deepseek_runner import DeepSeekRunner
 
         return DeepSeekRunner(args, model)
+    if model.model_style == LMStyle.QwenAPI:
+        from lcb_runner.runner.qwen_runner import QwenRunner
+
+        return QwenRunner(args, model)
     elif model.model_style in []:
         raise NotImplementedError(
             f"Runner for language model style {model.model_style} not implemented yet"