model-library 0.1.4__tar.gz → 0.1.6__tar.gz

{model_library-0.1.4 → model_library-0.1.6}/Makefile

@@ -16,8 +16,10 @@ help:
 	@echo "  make examples <model> Run all examples with specified model"
 	@echo "  make browse_models    Interactively browse models and their configurations"
 
+PYTHON_VERSION ?= 3.11
+
 install:
-	uv venv
+	uv venv --python $(PYTHON_VERSION)
 	uv sync --dev
 	@echo "🎉 Done! Run 'source .venv/bin/activate' to activate the environment locally."
 

{model_library-0.1.4 → model_library-0.1.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.4
+Version: 0.1.6
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT

{model_library-0.1.4 → model_library-0.1.6}/model_library/base/base.py

@@ -6,7 +6,6 @@ from abc import ABC, abstractmethod
 from collections.abc import Awaitable
 from pprint import pformat
 from typing import (
-    TYPE_CHECKING,
     Any,
     Callable,
     Literal,
@@ -43,9 +42,6 @@ from model_library.exceptions import (
 )
 from model_library.utils import truncate_str
 
-if TYPE_CHECKING:
-    from model_library.providers.openai import OpenAIModel
-
 PydanticT = TypeVar("PydanticT", bound=BaseModel)
 
 
@@ -66,7 +62,7 @@ class LLMConfig(BaseModel):
     top_p: float | None = None
     top_k: int | None = None
     reasoning: bool = False
-    reasoning_effort: str | None = None
+    reasoning_effort: str | bool | None = None
     supports_images: bool = False
     supports_files: bool = False
     supports_videos: bool = False
@@ -110,7 +106,7 @@ class LLM(ABC):
         self.top_k: int | None = config.top_k
 
         self.reasoning: bool = config.reasoning
-        self.reasoning_effort: str | None = config.reasoning_effort
+        self.reasoning_effort: str | bool | None = config.reasoning_effort
 
         self.supports_files: bool = config.supports_files
         self.supports_videos: bool = config.supports_videos
@@ -120,7 +116,7 @@ class LLM(ABC):
         self.supports_tools: bool = config.supports_tools
 
         self.native: bool = config.native
-        self.delegate: "OpenAIModel | None" = None
+        self.delegate: "LLM | None" = None
         self.batch: LLMBatchMixin | None = None
 
         if config.provider_config:
@@ -198,11 +194,14 @@ class LLM(ABC):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition] = [],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         if not self.delegate:
             raise Exception("Delegate not set")
-        return await self.delegate._query_impl(input, tools=tools, **kwargs)  # pyright: ignore[reportPrivateUsage]
+        return await self.delegate._query_impl(  # pyright: ignore[reportPrivateUsage]
+            input, tools=tools, query_logger=query_logger, **kwargs
+        )
 
     async def query(
         self,
@@ -213,6 +212,7 @@ class LLM(ABC):
         # for backwards compatibility
         files: list[FileInput] = [],
         images: list[FileInput] = [],
+        query_logger: logging.Logger | None = None,
         **kwargs: object,
     ) -> QueryResult:
         """
@@ -256,15 +256,18 @@ class LLM(ABC):
         input = [*history, *input]
 
         # unique logger for the query
-        query_id = uuid.uuid4().hex[:14]
-        query_logger = self.logger.getChild(f"query={query_id}")
+        if not query_logger:
+            query_id = uuid.uuid4().hex[:14]
+            query_logger = self.logger.getChild(f"query={query_id}")
 
         query_logger.info(
             "Query started:\n" + item_info + tool_info + f"--- kwargs: {short_kwargs}\n"
         )
 
         async def query_func() -> QueryResult:
-            return await self._query_impl(input, tools=tools, **kwargs)
+            return await self._query_impl(
+                input, tools=tools, query_logger=query_logger, **kwargs
+            )
 
         async def timed_query() -> tuple[QueryResult, float]:
             return await LLM.timer_wrapper(query_func)
@@ -361,7 +364,8 @@ class LLM(ABC):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
-        **kwargs: object,  # TODO: pass in query logger
+        query_logger: logging.Logger,
+        **kwargs: object,
     ) -> QueryResult:
         """
         Query the model with input

{model_library-0.1.4 → model_library-0.1.6}/model_library/base/delegate_only.py

@@ -1,4 +1,5 @@
 import io
+import logging
 from typing import Any, Literal, Sequence
 
 from typing_extensions import override
@@ -48,11 +49,14 @@ class DelegateOnly(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         assert self.delegate
 
-        return await self.delegate_query(input, tools=tools, **kwargs)
+        return await self.delegate_query(
+            input, tools=tools, query_logger=query_logger, **kwargs
+        )
 
     @override
     async def parse_input(

{model_library-0.1.4 → model_library-0.1.6}/model_library/base/output.py

@@ -9,9 +9,7 @@ from pydantic import BaseModel, Field, computed_field, field_validator
 from typing_extensions import override
 
 from model_library.base.input import InputItem, ToolCall
-from model_library.base.utils import (
-    sum_optional,
-)
+from model_library.base.utils import add_optional
 from model_library.utils import truncate_str
 
 
@@ -42,10 +40,14 @@ class QueryResultCost(BaseModel):
     reasoning: float | None = None
     cache_read: float | None = None
     cache_write: float | None = None
+    total_override: float | None = None
 
     @computed_field
     @property
     def total(self) -> float:
+        if self.total_override is not None:
+            return self.total_override
+
         return sum(
             filter(
                 None,
@@ -86,6 +88,16 @@ class QueryResultCost(BaseModel):
             )
         )
 
+    def __add__(self, other: "QueryResultCost") -> "QueryResultCost":
+        return QueryResultCost(
+            input=self.input + other.input,
+            output=self.output + other.output,
+            reasoning=add_optional(self.reasoning, other.reasoning),
+            cache_read=add_optional(self.cache_read, other.cache_read),
+            cache_write=add_optional(self.cache_write, other.cache_write),
+            total_override=add_optional(self.total_override, other.total_override),
+        )
+
     @override
     def __repr__(self):
         use_cents = self.total < 1
@@ -150,18 +162,20 @@ class QueryResultMetadata(BaseModel):
         return QueryResultMetadata(
             in_tokens=self.in_tokens + other.in_tokens,
             out_tokens=self.out_tokens + other.out_tokens,
-            reasoning_tokens=sum_optional(
-                self.reasoning_tokens, other.reasoning_tokens
+            reasoning_tokens=cast(
+                int | None, add_optional(self.reasoning_tokens, other.reasoning_tokens)
             ),
-            cache_read_tokens=sum_optional(
-                self.cache_read_tokens, other.cache_read_tokens
+            cache_read_tokens=cast(
+                int | None,
+                add_optional(self.cache_read_tokens, other.cache_read_tokens),
             ),
-            cache_write_tokens=sum_optional(
-                self.cache_write_tokens, other.cache_write_tokens
+            cache_write_tokens=cast(
+                int | None,
+                add_optional(self.cache_write_tokens, other.cache_write_tokens),
             ),
             duration_seconds=self.default_duration_seconds
             + other.default_duration_seconds,
-            cost=self.cost,
+            cost=cast(QueryResultCost | None, add_optional(self.cost, other.cost)),
         )
 
     @override

{model_library-0.1.4 → model_library-0.1.6}/model_library/base/utils.py

@@ -1,4 +1,4 @@
-from typing import Sequence, cast
+from typing import Sequence, TypeVar, cast
 
 from model_library.base.input import (
     FileBase,
@@ -8,17 +8,39 @@ from model_library.base.input import (
     ToolResult,
 )
 from model_library.utils import truncate_str
+from pydantic import BaseModel
 
+T = TypeVar("T", bound=BaseModel)
 
-def sum_optional(a: int | None, b: int | None) -> int | None:
-    """Sum two optional integers, returning None if both are None.
+
+def add_optional(
+    a: int | float | T | None, b: int | float | T | None
+) -> int | float | T | None:
+    """Add two optional objects, returning None if both are None.
 
     Preserves None to indicate "unknown/not provided" when both inputs are None,
-    otherwise treats None as 0 for summation.
+    otherwise returns the non-None value or their sum.
     """
     if a is None and b is None:
         return None
-    return (a or 0) + (b or 0)
+
+    if a is None or b is None:
+        return a or b
+
+    if isinstance(a, (int, float)) and isinstance(b, (int, float)):
+        return a + b
+
+    # NOTE: Ensure that the subtypes are the same so we can use the __add__ method just from one
+    if type(a) is type(b):
+        add_method = getattr(a, "__add__", None)
+        if add_method is not None:
+            return add_method(b)
+    else:
+        raise ValueError(
+            f"Cannot add {type(a)} and {type(b)} because they are not the same subclass"
+        )
+
+    return None
 
 
 def get_pretty_input_types(input: Sequence["InputItem"], verbose: bool = False) -> str:

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/all_models.json

@@ -1,4 +1,144 @@
 {
+    "minimax/MiniMax-M2.1": {
+        "company": "MiniMax",
+        "label": "MiniMax-M2.1",
+        "description": null,
+        "release_date": "2025-12-23",
+        "open_source": true,
+        "documentation_url": "https://platform.minimax.io/docs",
+        "properties": {
+            "context_window": 204800,
+            "max_tokens": 131000,
+            "training_cutoff": null,
+            "reasoning_model": true
+        },
+        "supports": {
+            "images": false,
+            "files": false,
+            "temperature": true,
+            "tools": true
+        },
+        "metadata": {
+            "deprecated": false,
+            "available_for_everyone": true,
+            "available_as_evaluator": false,
+            "ignored_for_cost": false
+        },
+        "provider_properties": {},
+        "costs_per_million_token": {
+            "input": 0.3,
+            "output": 1.2,
+            "cache": {
+                "read": 0.03,
+                "write": 0.375,
+                "write_markup": 1.0
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 1.0,
+            "top_p": 0.95
+        },
+        "provider_endpoint": "MiniMax-M2.1",
+        "provider_name": "minimax",
+        "full_key": "minimax/MiniMax-M2.1",
+        "slug": "minimax_MiniMax-M2.1"
+    },
+    "zai/glm-4.7": {
+        "company": "zAI",
+        "label": "GLM 4.7",
+        "description": "Latest model from ZAI",
+        "release_date": "2025-12-22",
+        "open_source": true,
+        "documentation_url": "https://docs.z.ai/",
+        "properties": {
+            "context_window": 200000,
+            "max_tokens": 128000,
+            "training_cutoff": null,
+            "reasoning_model": true
+        },
+        "supports": {
+            "images": false,
+            "files": false,
+            "temperature": true,
+            "tools": true
+        },
+        "metadata": {
+            "deprecated": false,
+            "available_for_everyone": true,
+            "available_as_evaluator": false,
+            "ignored_for_cost": false
+        },
+        "provider_properties": {},
+        "costs_per_million_token": {
+            "input": 0.6,
+            "output": 2.2,
+            "cache": {
+                "read": 0.11,
+                "read_discount": 1.0,
+                "write_markup": 1.0
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 1.0,
+            "top_p": 1.0
+        },
+        "provider_endpoint": "glm-4.7",
+        "provider_name": "zai",
+        "full_key": "zai/glm-4.7",
+        "slug": "zai_glm-4.7"
+    },
+    "google/gemini-3-flash-preview": {
+        "company": "Google",
+        "label": "Gemini 3 Flash (12/25)",
+        "description": "Google's newest budget workhorse model",
+        "release_date": "2025-12-17",
+        "open_source": false,
+        "documentation_url": "https://ai.google.dev/gemini-api/docs/models",
+        "properties": {
+            "context_window": 1048576,
+            "max_tokens": 65536,
+            "training_cutoff": null,
+            "reasoning_model": true
+        },
+        "supports": {
+            "images": true,
+            "videos": true,
+            "files": true,
+            "batch": true,
+            "temperature": true,
+            "tools": true
+        },
+        "metadata": {
+            "deprecated": false,
+            "available_for_everyone": true,
+            "available_as_evaluator": false,
+            "ignored_for_cost": false
+        },
+        "provider_properties": {},
+        "costs_per_million_token": {
+            "input": 0.5,
+            "output": 3.0,
+            "cache": {
+                "read_discount": 0.1,
+                "write_markup": 1.0
+            },
+            "batch": {
+                "input_discount": 0.5,
+                "output_discount": 0.5
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 1.0,
+            "reasoning_effort": "high"
+        },
+        "provider_endpoint": "gemini-3-flash-preview",
+        "provider_name": "google",
+        "full_key": "google/gemini-3-flash-preview",
+        "slug": "google_gemini-3-flash-preview"
+    },
     "openai/gpt-5.2-pro-2025-12-11": {
         "company": "OpenAI",
         "label": "GPT 5.2 Pro",
@@ -454,7 +594,8 @@
             }
         ],
         "default_parameters": {
-            "temperature": 1.0
+            "temperature": 1.0,
+            "reasoning_effort": "none"
         },
         "provider_endpoint": "deepseek-v3p2",
         "provider_name": "fireworks",
@@ -15428,7 +15569,7 @@
             "tools": false
         },
         "metadata": {
-            "deprecated": false,
+            "deprecated": true,
             "available_for_everyone": true,
             "available_as_evaluator": false,
             "ignored_for_cost": false

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/fireworks_models.yaml

@@ -150,6 +150,8 @@ deepseek-models:
       context_window: 160_000
       max_tokens: 20_480
       reasoning_model: false
+    default_parameters:
+      reasoning_effort: "none"
     costs_per_million_token:
       input: 0.56
       output: 1.68

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/google_models.yaml

@@ -54,6 +54,21 @@ gemini-3-models:
       temperature: 1
       reasoning_effort: "high"
 
+  google/gemini-3-flash-preview:
+    label: Gemini 3 Flash (12/25)
+    description: Google's newest budget workhorse model
+    release_date: 2025-12-17
+    properties:
+      context_window: 1048576
+      max_tokens: 65536
+      reasoning_model: true
+    costs_per_million_token:
+      input: 0.50
+      output: 3.00
+    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.

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/minimax_models.yaml

@@ -16,6 +16,24 @@ base-config:
 
 minimax-m2-models:
 
+  minimax/MiniMax-M2.1:
+    label: MiniMax-M2.1
+    release_date: 2025-12-23
+    properties:
+      context_window: 204_800
+      max_tokens: 131_000
+      reasoning_model: true
+      training_cutoff: null
+    default_parameters:
+      temperature: 1.0
+      top_p: 0.95
+    costs_per_million_token:
+      input: 0.30
+      output: 1.20
+      cache:
+        read: 0.03
+        write: 0.375
+
   minimax/MiniMax-M2:
     label: MiniMax-M2
     description: MiniMax-M2 is a cost-efficient open-source model optimized for agentic applications and coding in particular.

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/perplexity_models.yaml

@@ -46,6 +46,8 @@ perplexity-models:
     label: Sonar Reasoning
     description: Reasoning-focused search model that exposes intermediate thinking for step-by-step answers.
     documentation_url: https://docs.perplexity.ai/models/models/sonar-reasoning
+    metadata:
+      deprecated: true
     properties:
       context_window: 128000
       reasoning_model: true

{model_library-0.1.4 → model_library-0.1.6}/model_library/config/zai_models.yaml

@@ -18,6 +18,20 @@ base-config:
       write_markup: 1
 
 zai-models:
+  zai/glm-4.7:
+    label: GLM 4.7
+    description: "Latest model from ZAI"
+    release_date: 2025-12-22
+    properties:
+      context_window: 200_000
+      max_tokens: 128_000
+    costs_per_million_token:
+      input: 0.6
+      output: 2.2
+      cache:
+        read: 0.11
+    default_parameters:
+      temperature: 1
   zai/glm-4.5:
     label: GLM 4.5
     description: "z.AI old model"

{model_library-0.1.4 → model_library-0.1.6}/model_library/exceptions.py

@@ -183,8 +183,8 @@ RETRIABLE_EXCEPTION_CODES = [
     "server_error",
     "overloaded",
     "throttling",  # AWS throttling errors
-    "throttlingexception",  # AWS throttling errors
     "internal server error",
+    "InternalServerError",
 ]
 
 

{model_library-0.1.4 → model_library-0.1.6}/model_library/providers/ai21labs.py

@@ -1,4 +1,5 @@
 import io
+import logging
 from typing import Any, Literal, Sequence
 
 from ai21 import AsyncAI21Client
@@ -137,6 +138,7 @@ class AI21LabsModel(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         messages: list[ChatMessage] = []

{model_library-0.1.4 → model_library-0.1.6}/model_library/providers/amazon.py

@@ -3,6 +3,7 @@ import asyncio
 import base64
 import io
 import json
+import logging
 from typing import Any, Literal, Sequence, cast
 
 import boto3
@@ -337,6 +338,7 @@ class AmazonModel(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         body = await self.build_body(input, tools=tools, **kwargs)

{model_library-0.1.4 → model_library-0.1.6}/model_library/providers/anthropic.py

@@ -1,4 +1,5 @@
 import io
+import logging
 from typing import Any, Literal, Sequence, cast
 
 from anthropic import AsyncAnthropic
@@ -249,6 +250,8 @@ class AnthropicModel(LLM):
 
     @override
     def get_client(self) -> AsyncAnthropic:
+        if self._delegate_client:
+            return self._delegate_client
         if not AnthropicModel._client:
             headers: dict[str, str] = {}
             AnthropicModel._client = AsyncAnthropic(
@@ -262,16 +265,20 @@ class AnthropicModel(LLM):
     def __init__(
         self,
         model_name: str,
-        provider: Literal["anthropic"] = "anthropic",
+        provider: str = "anthropic",
         *,
         config: LLMConfig | None = None,
+        custom_client: AsyncAnthropic | None = None,
     ):
         super().__init__(model_name, provider, config=config)
 
+        # allow custom client to act as delegate (native)
+        self._delegate_client: AsyncAnthropic | None = custom_client
+
         # https://docs.anthropic.com/en/api/openai-sdk
-        self.delegate: OpenAIModel | None = (
+        self.delegate = (
             None
-            if self.native
+            if self.native or custom_client
             else OpenAIModel(
                 model_name=self.model_name,
                 provider=provider,
@@ -285,7 +292,10 @@ class AnthropicModel(LLM):
         )
 
         # Initialize batch support if enabled
-        self.supports_batch: bool = self.supports_batch and self.native
+        # Disable batch when using custom_client (similar to OpenAI)
+        self.supports_batch: bool = (
+            self.supports_batch and self.native and not custom_client
+        )
         self.batch: LLMBatchMixin | None = (
             AnthropicBatchMixin(self) if self.supports_batch else None
         )
@@ -555,20 +565,36 @@ class AnthropicModel(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         if self.delegate:
-            return await self.delegate_query(input, tools=tools, **kwargs)
+            return await self.delegate_query(
+                input, tools=tools, query_logger=query_logger, **kwargs
+            )
 
         body = await self.create_body(input, tools=tools, **kwargs)
 
-        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")
+        client = self.get_client()
 
-        async with self.get_client().beta.messages.stream(
-            **body,
-            betas=betas,
+        # only send betas for the official Anthropic endpoint
+        is_anthropic_endpoint = self._delegate_client is None
+        if not is_anthropic_endpoint:
+            client_base_url = getattr(client, "_base_url", None) or getattr(
+                client, "base_url", None
+            )
+            if client_base_url:
+                is_anthropic_endpoint = "api.anthropic.com" in str(client_base_url)
+
+        stream_kwargs = {**body}
+        if is_anthropic_endpoint:
+            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")
+            stream_kwargs["betas"] = betas
+
+        async with client.beta.messages.stream(
+            **stream_kwargs,
         ) as stream:  # pyright: ignore[reportAny]
             message = await stream.get_final_message()
         self.logger.info(f"Anthropic Response finished: {message.id}")