model-library 0.1.2__py3-none-any.whl → 0.1.3__py3-none-any.whl

model_library/config/anthropic_models.yaml

@@ -27,6 +27,28 @@ claude-4-models:
     class_properties:
       supports_batch_requests: true
 
+  anthropic/claude-opus-4-5-20251101:
+    label: Claude Opus 4.5 (Nonthinking)
+    release_date: 2025-11-24
+    properties:
+      context_window: 200_000
+      max_token_output: 64_000
+      extending_thinking: 64_000
+    class_properties:
+      available_for_everyone: false
+    default_parameters:
+      max_output_tokens: 64_000
+    costs_per_million_token:
+      input: 15.0
+      output: 75.0
+      cache:
+        read: 1.5
+        write: 18.75
+    alternative_keys:
+      - anthropic/claude-opus-4-5-20251101-thinking:
+          properties:
+            reasoning_model: true
+
   anthropic/claude-opus-4-1-20250805:
     label: Claude Opus 4.1 (Nonthinking)
     description: Advanced model for specialized complex
@@ -197,11 +219,12 @@ claude-3-5-models:
     alternative_keys:
       - anthropic/claude-3-5-sonnet-latest
       - anthropic/claude-3.5-sonnet-latest
-  
+
   anthropic/claude-3-5-sonnet-20240620:
     label: Claude 3.5 Sonnet
     release_date: 2024-06-20
-    description: Claude Sonnet 3.5 (June 2024) variant for code and content generation,
+    description:
+      Claude Sonnet 3.5 (June 2024) variant for code and content generation,
       multilingual and vision-capable, deprecated.
     class_properties:
       deprecated: true
@@ -339,7 +362,7 @@ claude-2-models:
     costs_per_million_token:
       input: 8.0
       output: 24.0
-  
+
   anthropic/claude-1.3:
     label: Claude 1.3
     release_date: null

model_library/config/google_models.yaml

@@ -31,6 +31,55 @@ gemma-models:
       input: 0.00
       output: 0.00
 
+
+gemini-3-models:
+  base-config:
+    properties:
+      context_window: 1_048_576
+      max_token_output: 8_192
+      training_cutoff: "2025-01"
+    class_properties:
+      supports_images: true
+      supports_files: true
+      supports_videos: true
+      supports_tools: true
+      supports_batch_requests: true
+      supports_temperature: true
+    costs_per_million_token:
+      cache:
+        read_discount: 0.1
+    default_parameters:
+      temperature: 1
+      reasoning_effort: "high"
+
+  google/gemini-3-pro-preview:
+    label: Gemini 3 Pro (11/25)
+    description: Gemini 3 Pro, Google's most powerful model.
+    release_date: 2025-11-18
+    properties:
+      context_window: 1048576
+      max_token_output: 65536
+      training_cutoff: "2025-01"
+      reasoning_model: true
+    class_properties:
+      supports_images: true
+      supports_files: true
+      supports_videos: true
+      supports_tools: true
+      supports_batch_requests: true
+      supports_temperature: true
+    costs_per_million_token:
+      input: 2
+      output: 12
+      cache:
+        read: 0.20
+      context:
+        threshold: 200_000
+        input: 2.5
+        output: 15.0
+        cache:
+          read: 0.40
+
 gemini-2.5-models:
   base-config:
     properties:

model_library/config/openai_models.yaml

@@ -31,7 +31,6 @@ gpt-5-models:
       supports_temperature: false
       supports_files: true
       supports_tools: true
-  
   openai/gpt-5.1-codex:
     label: GPT 5.1 Codex
     documentation_url: https://platform.openai.com/docs/models/gpt-5.1-codex
@@ -43,7 +42,6 @@ gpt-5-models:
       cache:
         read: 0.125
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
 
   openai/gpt-5.1-codex-mini:
@@ -57,7 +55,6 @@ gpt-5-models:
       cache:
         read: 0.025
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
 
   openai/gpt-5-codex:
@@ -76,10 +73,8 @@ gpt-5-models:
       available_as_evaluator: true
       supports_images: true
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
 
-  
   openai/gpt-5.1-2025-11-13:
     label: GPT 5.1
     documentation_url: https://platform.openai.com/docs/models/gpt-5.1
@@ -96,7 +91,6 @@ gpt-5-models:
       available_as_evaluator: true
       supports_images: true
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
 
   openai/gpt-5-2025-08-07:
@@ -115,7 +109,6 @@ gpt-5-models:
       available_as_evaluator: true
       supports_images: true
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
     alternative_keys:
       - azure/gpt-5-2025-08-07
@@ -133,7 +126,6 @@ gpt-5-models:
     properties:
       training_cutoff: "2024-05"
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
     class_properties:
       supports_images: true
@@ -153,7 +145,6 @@ gpt-5-models:
     properties:
       training_cutoff: "2024-05"
     default_parameters:
-      temperature: 1
       max_output_tokens: 128_000
     class_properties:
       supports_images: true

model_library/config/together_models.yaml

@@ -28,6 +28,7 @@ kimi-models:
       supports_temperature: true
     default_parameters:
       temperature: 0.3
+      max_output_tokens: 16_384
 
   together/moonshotai/Kimi-K2-Instruct:
     label: Kimi K2 Instruct

model_library/config/xai_models.yaml

@@ -8,7 +8,7 @@ base-config:
     supports_files: false
     available_for_everyone: true
     ignored_for_cost: false
-    supports_tools: false
+    supports_tools: true
   properties:
     reasoning_model: false
 
@@ -33,7 +33,6 @@ xai-models:
       reasoning_model: true
     class_properties:
       supports_images: false
-      supports_tools: true
     costs_per_million_token:
       input: 0.20
       output: 1.50
@@ -81,6 +80,68 @@ xai-models:
       - grok/grok-4-fast
       - grok/grok-4-fast-reasoning-latest
 
+  grok/grok-4-1-fast-reasoning:
+    label: Grok 4.1 Fast (Reasoning)
+    description: ""
+    release_date: 2025-10-19
+    open_source: false
+    class_properties:
+      supports_images: true
+      available_as_evaluator: true
+      supports_metadata: true
+      supports_files: false
+      available_for_everyone: true
+      ignored_for_cost: false
+    properties:
+      context_window: 2_000_000
+      max_token_output: 2_000_000 # from openrouter
+      training_cutoff: null
+      reasoning_model: true
+    documentation_url: ""
+    costs_per_million_token:
+      input: 0.20
+      output: 0.5
+      cache:
+        read: 0.05
+      context:
+        threshold: 128_000
+        input: 0.4
+        output: 1.0
+    default_parameters:
+      temperature: 0.7
+      max_output_tokens: 128000
+
+  grok/grok-4-1-fast-non-reasoning:
+    label: Grok 4.1 Fast Non-Reasoning
+    description: ""
+    release_date: 2025-10-19
+    open_source: false
+    class_properties:
+      supports_images: true
+      available_as_evaluator: true
+      supports_metadata: true
+      supports_files: false
+      available_for_everyone: true
+      ignored_for_cost: false
+    properties:
+      context_window: 2_000_000
+      max_token_output: 2_000_000 # from openrouter
+      training_cutoff: null
+      reasoning_model: false
+    documentation_url: ""
+    costs_per_million_token:
+      input: 0.20
+      output: 0.5
+      cache:
+        read: 0.05
+      context:
+        threshold: 128_000
+        input: 0.4
+        output: 1.0
+    default_parameters:
+      temperature: 0.7
+      max_output_tokens: 128000
+
   grok/grok-4-fast-non-reasoning:
     label: Grok 4 Fast (Non-Reasoning)
     description: Cost-efficient model focused on speed and efficiency for straightforward tasks like summarization or classification without deep logical processing. Unified architecture with reasoning variant, steered via system prompts.
@@ -121,7 +182,6 @@ xai-models:
     class_properties:
       supports_images: true
       available_for_everyone: false
-      supports_tools: true
     properties:
       context_window: 256_000
       max_token_output: 128_000

model_library/exceptions.py

@@ -5,6 +5,7 @@ from typing import Any, Callable
 
 import backoff
 from ai21 import TooManyRequestsError as AI21RateLimitError
+from anthropic import InternalServerError
 from anthropic import RateLimitError as AnthropicRateLimitError
 from backoff._typing import Details
 from httpcore import ReadError as HTTPCoreReadError
@@ -166,6 +167,7 @@ RETRIABLE_EXCEPTIONS = [
     OpenAIUnprocessableEntityError,
     OpenAIAPIConnectionError,
     AnthropicRateLimitError,
+    InternalServerError,
     AI21RateLimitError,
     RemoteProtocolError,  # httpx connection closing when running models from sdk
     HTTPXReadError,
@@ -191,6 +193,7 @@ RETRIABLE_EXCEPTION_CODES = [
     "overloaded",
     "throttling",  # AWS throttling errors
     "throttlingexception",  # AWS throttling errors
+    "internal server error",
 ]
 
 
@@ -239,8 +242,9 @@ def retry_llm_call(
     logger: logging.Logger,
     max_tries: int = RETRY_MAX_TRIES,
     max_time: float | None = None,
-    backoff_callback: Callable[[int, Exception | None, float, float], None]
-    | None = None,
+    backoff_callback: (
+        Callable[[int, Exception | None, float, float], None] | None
+    ) = None,
 ):
     def on_backoff(details: Details):
         exception = details.get("exception")

model_library/file_utils.py

@@ -56,7 +56,7 @@ def concat_images(
         new_width = int(combined_image.width * scale_factor)
         new_height = int(combined_image.height * scale_factor)
 
-        combined_image = combined_image.resize(
+        combined_image = combined_image.resize(  # type: ignore
             (new_width, new_height), Image.Resampling.LANCZOS
         )
 

model_library/providers/anthropic.py

@@ -562,12 +562,8 @@ class AnthropicModel(LLM):
 
         body = await self.create_body(input, tools=tools, **kwargs)
 
-        betas = [
-            "files-api-2025-04-14",
-            "interleaved-thinking-2025-05-14",
-        ]
-
-        if "claude-sonnet-4-5" in self.model_name:
+        betas = ["files-api-2025-04-14", "interleaved-thinking-2025-05-14"]
+        if "sonnet-4-5" in self.model_name:
             betas.append("context-1m-2025-08-07")
 
         async with self.get_client().beta.messages.stream(

model_library/providers/google/google.py

@@ -2,8 +2,6 @@ import base64
 import io
 from typing import Any, Literal, Sequence, cast
 
-from typing_extensions import override
-
 from google.genai import Client
 from google.genai import errors as genai_errors
 from google.genai.types import (
@@ -18,10 +16,14 @@ from google.genai.types import (
     Part,
     SafetySetting,
     ThinkingConfig,
+    ThinkingLevel,
     Tool,
     ToolListUnion,
     UploadFileConfig,
+    FinishReason,
 )
+from typing_extensions import override
+
 from model_library import model_library_settings
 from model_library.base import (
     LLM,
@@ -119,15 +121,6 @@ class GoogleModel(LLM):
     ):
         super().__init__(model_name, provider, config=config)
 
-        # thinking tag
-        if self.model_name.endswith("-thinking"):
-            original_name = self.model_name
-            self.model_name = self.model_name.replace("-thinking", "")
-            self.reasoning = True
-            self.logger.info(
-                f"Enabled thinking mode for {original_name} -> {self.model_name}"
-            )
-
         if self.provider_config.use_vertex:
             self.supports_batch = False
 
@@ -261,14 +254,12 @@ class GoogleModel(LLM):
         bytes: io.BytesIO,
         type: Literal["image", "file"] = "file",
     ) -> FileWithId:
-        if not self.supports_batch:
-            if self.provider_config.use_vertex:
-                raise Exception(
-                    "Vertex AI does not support file uploads. "
-                    "use FileWithBase64 to pass files as inline data"
-                    "or use genai for file uploads"
-                )
-            raise Exception("Model does not support batching")
+        if self.provider_config.use_vertex:
+            raise Exception(
+                "Vertex AI does not support file uploads. "
+                "use FileWithBase64 to pass files as inline data"
+                "or use genai for file uploads"
+            )
 
         mime = f"image/{mime}" if type == "image" else mime  # TODO:
         response: File = self.client.files.upload(
@@ -294,7 +285,6 @@ class GoogleModel(LLM):
         tools: list[ToolDefinition],
         **kwargs: object,
     ) -> dict[str, Any]:
-        self.logger.debug(f"Creating request body for {self.model_name}")
         generation_config = GenerateContentConfig(
             max_output_tokens=self.max_tokens,
         )
@@ -310,13 +300,15 @@ class GoogleModel(LLM):
         if system_prompt and isinstance(system_prompt, str) and system_prompt.strip():
             generation_config.system_instruction = str(system_prompt)
 
-        if "gemini-2.5" in self.model_name and self.reasoning:
-            generation_config.thinking_config = ThinkingConfig(
-                thinking_budget=cast(
+        if self.reasoning:
+            reasoning_config = ThinkingConfig(include_thoughts=True)
+            if self.reasoning_effort:
+                reasoning_config.thinking_level = ThinkingLevel(self.reasoning_effort)
+            else:
+                reasoning_config.thinking_budget = cast(
                     int, kwargs.pop("thinking_budget", self.DEFAULT_THINKING_BUDGET)
-                ),
-                include_thoughts=True,
-            )
+                )
+            generation_config.thinking_config = reasoning_config
 
         if tools:
             generation_config.tools = cast(ToolListUnion, await self.parse_tools(tools))
@@ -343,17 +335,20 @@ class GoogleModel(LLM):
         text: str = ""
         reasoning: str = ""
         tool_calls: list[ToolCall] = []
-        last_content: Content | None = None
 
         metadata: GenerateContentResponseUsageMetadata | None = None
 
         stream = await self.client.aio.models.generate_content_stream(**body)
+        contents: list[Content | None] = []
+        finish_reason: FinishReason | None = None
+
         async for chunk in stream:
             candidates = chunk.candidates
             if not candidates:
                 continue
 
             content = candidates[0].content
+
             if content and content.parts:
                 for part in content.parts:
                     if part.function_call:
@@ -378,14 +373,24 @@ class GoogleModel(LLM):
 
             if chunk.usage_metadata:
                 metadata = chunk.usage_metadata
-            last_content = content
+            if content:
+                contents.append(content)
+            if candidates[0].finish_reason:
+                finish_reason = candidates[0].finish_reason
+
+        if finish_reason != FinishReason.STOP:
+            self.logger.error(f"Unexpected finish reason: {finish_reason}")
+
+        if not text and not reasoning and not tool_calls:
+            raise ModelNoOutputError("Model returned empty response")
 
         result = QueryResult(
             output_text=text,
             reasoning=reasoning,
-            history=[*input, last_content],
+            history=[*input, *contents],
             tool_calls=tool_calls,
         )
+
         if metadata:
             # see _calculate_cost
             cache_read_tokens = metadata.cached_content_token_count or 0
@@ -446,6 +451,7 @@ class GoogleModel(LLM):
                 "response_mime_type": "application/json",
             }
         )
+
         body["config"] = config
 
         # Make the request with retry wrapper

model_library/providers/openai.py

@@ -5,7 +5,11 @@ import json
 from typing import Any, Literal, Sequence, cast
 
 from openai import APIConnectionError, AsyncOpenAI
-from openai.types.chat import ChatCompletionMessage, ChatCompletionMessageToolCall
+from openai.types.chat import (
+    ChatCompletionMessage,
+    ChatCompletionMessageToolCall,
+    ChatCompletionMessageToolCallUnion,
+)
 from openai.types.chat.chat_completion_message_tool_call import Function
 from openai.types.chat.chat_completion_tool_param import ChatCompletionToolParam
 from openai.types.create_embedding_response import CreateEmbeddingResponse
@@ -617,7 +621,9 @@ class OpenAIModel(LLM):
         final_message = ChatCompletionMessage(
             role="assistant",
             content=output_text if output_text else None,
-            tool_calls=raw_tool_calls if raw_tool_calls else None,
+            tool_calls=cast(list[ChatCompletionMessageToolCallUnion], raw_tool_calls)
+            if raw_tool_calls
+            else None,
         )
         if hasattr(final_message, "reasoning_content") and reasoning_text:
             setattr(final_message, "reasoning_content", reasoning_text)