model-library 0.1.7__py3-none-any.whl → 0.1.9__py3-none-any.whl

model_library/config/mistral_models.yaml

@@ -38,6 +38,8 @@ magistral-models:
     costs_per_million_token:
       input: 2
       output: 5
+    metadata:
+      deprecated: true
 
   mistralai/magistral-medium-2509:
     label: Magistral Medium 1.2 (09/2025)

model_library/config/openai_models.yaml

@@ -35,6 +35,19 @@ gpt-5-models:
       training_cutoff: "2024-09"
       reasoning_model: true
 
+  openai/gpt-5.2-codex:
+    label: GPT 5.2 Codex
+    documentation_url: https://platform.openai.com/docs/models/gpt-5.2-codex
+    description: GPT 5.2 optimized for code
+    release_date: 2025-12-11
+    properties:
+      context_window: 400_000
+    costs_per_million_token:
+      input: 1.75
+      output: 14
+      cache:
+        read: 0.175
+
   openai/gpt-5.2-2025-12-11:
     label: GPT 5.2
     documentation_url: https://platform.openai.com/docs/models/gpt-5.2
@@ -70,7 +83,7 @@ gpt-5-models:
   openai/gpt-5.1-codex-max:
     label: GPT 5.1 Codex Max
     release_date: 2025-12-04
-    description: OpenAI's frontier agentic coding model. Good at long-running coding tasks.
+    description: GPT 5.1 optimized for code
     costs_per_million_token:
       input: 1.25
       output: 10.0
@@ -79,7 +92,7 @@ gpt-5-models:
   openai/gpt-5.1-codex:
     label: GPT 5.1 Codex
     documentation_url: https://platform.openai.com/docs/models/gpt-5.1-codex
-    description: OpenAI's latest coding model
+    description: GPT 5.1 optimized for code
     release_date: 2025-11-13
     costs_per_million_token:
       input: 1.25
@@ -841,24 +854,3 @@ gpt-3.5-models:
       input: 1.5
       output: 2.0
     documentation_url: https://platform.openai.com/docs/models/gpt-3.5-turbo-instruct
-
-databricks-models:
-  base-config:
-    company: Databricks
-
-  databricks/databricks-dbrx-instruct:
-    label: DBRX Instruct
-    description: Databricks Instruct model.
-    release_date: 2024-03-27
-    properties:
-      context_window: 32_768
-      max_tokens: 4_096
-      training_cutoff: "2023-12"
-    metadata:
-      available_for_everyone: false
-      deprecated: true
-    costs_per_million_token:
-      input: 2.25
-      output: 6.75
-    alternative_keys:
-      - databricks/dbrx-instruct

model_library/config/together_models.yaml

@@ -45,6 +45,8 @@ kimi-models:
     costs_per_million_token:
       input: 1.00
       output: 3.00
+    metadata:
+      deprecated: true
 
 # Meta Llama Models
 llama-4-models:

model_library/config/xiaomi_models.yaml

@@ -0,0 +1,43 @@
+base-config:
+  company: Xiaomi
+  open_source: true
+  documentation_url: https://platform.xiaomimimo.com/#/docs/
+  supports:
+    images: false
+    files: false
+    tools: true
+  metadata:
+    available_as_evaluator: false
+    available_for_everyone: true
+    ignored_for_cost: false
+  properties:
+    training_cutoff: December 2024
+
+xiaomi-models:
+  base-config:
+    properties:
+      context_window: 256000
+    supports:
+      temperature: true
+      top_p: true
+    default_parameters:
+      temperature: 0.3
+      top_p: 0.95
+
+  xiaomi/mimo-v2-flash:
+    label: MiMo V2 Flash
+    description:
+      MiMo V2 Flash is Xiaomi's Mixture-of-Experts (MoE) language model with
+      309B total parameters and 15B active parameters. Designed for high-speed
+      reasoning and agentic workflows, it utilizes a novel hybrid attention
+      architecture and Multi-Token Prediction (MTP) to achieve state-of-the-art
+      performance while significantly reducing inference costs.
+    release_date: 2025-12-17
+    properties:
+      context_window: 256000
+      max_tokens: 64000
+    costs_per_million_token:
+      input: 0.10
+      output: 0.30
+      cache:
+        read: 0.01

model_library/config/zai_models.yaml

@@ -18,6 +18,25 @@ base-config:
       write_markup: 1
 
 zai-models:
+  zai/glm-4.7-flashx:
+    label: GLM 4.7 Flash
+    description: "z.AI lightweight fast model"
+    release_date: 2026-01-19
+    properties:
+      context_window: 200_000
+      max_tokens: 128_000
+    costs_per_million_token:
+      input: 0.07
+      output: 0.4
+      cache:
+        read: 0.01
+    default_parameters:
+      # from https://huggingface.co/zai-org/GLM-4.7-Flash
+      temperature: 1
+      top_p: 0.95
+    provider_properties:
+      clear_thinking: false
+  
   zai/glm-4.7:
     label: GLM 4.7
     description: "Latest model from ZAI"
@@ -32,6 +51,9 @@ zai-models:
         read: 0.11
     default_parameters:
       temperature: 1
+    alternative_keys:
+      - vercel/zai/glm-4.7
+
   zai/glm-4.5:
     label: GLM 4.5
     description: "z.AI old model"
@@ -46,9 +68,11 @@ zai-models:
         read: 0.11
     alternative_keys:
       - fireworks/glm-4p5:
-        costs_per_million_token:
-          input: 0.55
-          output: 2.19
+          metadata:
+            deprecated: true
+          costs_per_million_token:
+            input: 0.55
+            output: 2.19
 
   zai/glm-4.5-air:
     label: GLM 4.5 Air

model_library/exceptions.py

@@ -1,13 +1,9 @@
-import logging
-import random
 import re
-from typing import Any, Callable
+from typing import Any
 
-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
 from httpx import ConnectError as HTTPXConnectError
 from httpx import ReadError as HTTPXReadError
@@ -75,12 +71,14 @@ CONTEXT_WINDOW_PATTERN = re.compile(
     r"maximum context length is \d+ tokens|"
     r"context length is \d+ tokens|"
     r"exceed.* context (limit|window|length)|"
+    r"context window exceeds|"
     r"exceeds maximum length|"
     r"too long.*tokens.*maximum|"
     r"too large for model with \d+ maximum context length|"
     r"longer than the model's context length|"
     r"too many tokens.*size limit exceeded|"
     r"prompt is too long|"
+    r"maximum prompt length|"
     r"input length should be|"
     r"sent message larger than max|"
     r"input tokens exceeded|"
@@ -222,75 +220,3 @@ def exception_message(exception: Exception | Any) -> str:
         if str(exception)
         else type(exception).__name__
     )
-
-
-RETRY_MAX_TRIES: int = 20
-RETRY_INITIAL: float = 10.0
-RETRY_EXPO: float = 1.4
-RETRY_MAX_BACKOFF_WAIT: float = 240.0  # 4 minutes (more with jitter)
-
-
-def jitter(wait: float) -> float:
-    """
-    Increase or decrease the wait time by up to 20%.
-    """
-    jitter_fraction = 0.2
-    min_wait = wait * (1 - jitter_fraction)
-    max_wait = wait * (1 + jitter_fraction)
-    return random.uniform(min_wait, max_wait)
-
-
-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,
-):
-    def on_backoff(details: Details):
-        exception = details.get("exception")
-        tries = details.get("tries", 0)
-        elapsed = details.get("elapsed", 0.0)
-        wait = details.get("wait", 0.0)
-
-        logger.warning(
-            f"[Retrying] Exception: {exception_message(exception)} | Attempt: {tries} | "
-            + f"Elapsed: {elapsed:.1f}s | Next wait: {wait:.1f}s"
-        )
-
-        if backoff_callback:
-            backoff_callback(tries, exception, elapsed, wait)
-
-    def giveup(e: Exception) -> bool:
-        return not is_retriable_error(e)
-
-    def on_giveup(details: Details) -> None:
-        exception: Exception | None = details.get("exception", None)
-        if not exception:
-            return
-
-        logger.error(
-            f"Giving up after retries. Final exception: {exception_message(exception)}"
-        )
-
-        if is_context_window_error(exception):
-            message = exception.args[0] if exception.args else str(exception)
-            raise MaxContextWindowExceededError(message)
-
-        raise exception
-
-    return backoff.on_exception(
-        wait_gen=lambda: backoff.expo(
-            base=RETRY_EXPO,
-            factor=RETRY_INITIAL,
-            max_value=RETRY_MAX_BACKOFF_WAIT,
-        ),
-        exception=Exception,
-        max_tries=max_tries,
-        max_time=max_time,
-        giveup=giveup,
-        on_backoff=on_backoff,
-        on_giveup=on_giveup,
-        jitter=jitter,
-    )

model_library/providers/ai21labs.py

@@ -16,13 +16,13 @@ from model_library.base import (
     LLMConfig,
     QueryResult,
     QueryResultMetadata,
+    RawResponse,
     TextInput,
     ToolBody,
     ToolCall,
     ToolDefinition,
     ToolResult,
 )
-from model_library.base.input import RawResponse
 from model_library.exceptions import (
     BadInputError,
     MaxOutputTokensExceededError,
@@ -34,17 +34,21 @@ from model_library.utils import default_httpx_client
 
 @register_provider("ai21labs")
 class AI21LabsModel(LLM):
-    _client: AsyncAI21Client | None = None
+    @override
+    def _get_default_api_key(self) -> str:
+        return model_library_settings.AI21LABS_API_KEY
 
     @override
-    def get_client(self) -> AsyncAI21Client:
-        if not AI21LabsModel._client:
-            AI21LabsModel._client = AsyncAI21Client(
-                api_key=model_library_settings.AI21LABS_API_KEY,
+    def get_client(self, api_key: str | None = None) -> AsyncAI21Client:
+        if not self.has_client():
+            assert api_key
+            client = AsyncAI21Client(
+                api_key=api_key,
                 http_client=default_httpx_client(),
-                num_retries=1,
+                num_retries=3,
             )
-        return AI21LabsModel._client
+            self.assign_client(client)
+        return super().get_client()
 
     def __init__(
         self,

model_library/providers/alibaba.py

@@ -1,17 +1,17 @@
-from typing import Literal
+from typing import Any, Literal
 
+from pydantic import SecretStr
 from typing_extensions import override
 
 from model_library import model_library_settings
 from model_library.base import (
+    DelegateConfig,
     DelegateOnly,
     LLMConfig,
     QueryResultCost,
     QueryResultMetadata,
 )
-from model_library.providers.openai import OpenAIModel
 from model_library.register_models import register_provider
-from model_library.utils import create_openai_client_with_defaults
 
 
 @register_provider("alibaba")
@@ -26,17 +26,26 @@ class AlibabaModel(DelegateOnly):
         super().__init__(model_name, provider, config=config)
 
         # https://www.alibabacloud.com/help/en/model-studio/first-api-call-to-qwen
-        self.delegate = OpenAIModel(
-            model_name=self.model_name,
-            provider=self.provider,
+        self.init_delegate(
             config=config,
-            custom_client=create_openai_client_with_defaults(
-                api_key=model_library_settings.DASHSCOPE_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
+                api_key=SecretStr(model_library_settings.DASHSCOPE_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )
 
+    @override
+    def _get_extra_body(self) -> dict[str, Any]:
+        """Build extra body parameters for Qwen-specific features."""
+        extra: dict[str, Any] = {}
+        # Enable thinking mode for Qwen3 reasoning models
+        # https://www.alibabacloud.com/help/en/model-studio/use-qwen-by-calling-api
+        if self.reasoning:
+            extra["enable_thinking"] = True
+        return extra
+
     @override
     async def _calculate_cost(
         self,

model_library/providers/amazon.py

@@ -11,6 +11,7 @@ import botocore
 from botocore.client import BaseClient
 from typing_extensions import override
 
+from model_library import model_library_settings
 from model_library.base import (
     LLM,
     FileBase,
@@ -41,20 +42,46 @@ from model_library.register_models import register_provider
 @register_provider("amazon")
 @register_provider("bedrock")
 class AmazonModel(LLM):
-    _client: BaseClient | None = None
+    @override
+    def _get_default_api_key(self) -> str:
+        if getattr(model_library_settings, "AWS_ACCESS_KEY_ID", None):
+            return json.dumps(
+                {
+                    "AWS_ACCESS_KEY_ID": model_library_settings.AWS_ACCESS_KEY_ID,
+                    "AWS_SECRET_ACCESS_KEY": model_library_settings.AWS_SECRET_ACCESS_KEY,
+                    "AWS_DEFAULT_REGION": model_library_settings.AWS_DEFAULT_REGION,
+                }
+            )
+        return "using-environment"
 
     @override
-    def get_client(self) -> BaseClient:
-        if not AmazonModel._client:
-            AmazonModel._client = cast(
-                BaseClient,
-                boto3.client(
-                    "bedrock-runtime",
-                    # default connection pool is 10
-                    config=botocore.config.Config(max_pool_connections=1000),  # pyright: ignore[reportAttributeAccessIssue]
-                ),
-            )  # pyright: ignore[reportUnknownMemberType]
-        return AmazonModel._client
+    def get_client(self, api_key: str | None = None) -> BaseClient:
+        if not self.has_client():
+            assert api_key
+            if api_key != "using-environment":
+                creds = json.loads(api_key)
+                client = cast(
+                    BaseClient,
+                    boto3.client(
+                        "bedrock-runtime",
+                        aws_access_key_id=creds["AWS_ACCESS_KEY_ID"],
+                        aws_secret_access_key=creds["AWS_SECRET_ACCESS_KEY"],
+                        region_name=creds["AWS_DEFAULT_REGION"],
+                        config=botocore.config.Config(max_pool_connections=1000),  # pyright: ignore[reportAttributeAccessIssue]
+                    ),
+                )
+            else:
+                client = cast(
+                    BaseClient,
+                    boto3.client(
+                        "bedrock-runtime",
+                        # default connection pool is 10
+                        config=botocore.config.Config(max_pool_connections=1000),  # pyright: ignore[reportAttributeAccessIssue]
+                    ),
+                )
+
+            self.assign_client(client)
+        return super().get_client()
 
     def __init__(
         self,
@@ -70,6 +97,11 @@ class AmazonModel(LLM):
         )  # supported but no access yet
         self.supports_tool_cache = self.supports_cache and "claude" in self.model_name
 
+        if config and config.custom_api_key:
+            raise Exception(
+                "custom_api_key is not currently supported for Amazon models"
+            )
+
     cache_control = {"type": "default"}
 
     async def get_tool_call_ids(self, input: Sequence[InputItem]) -> list[str]:
@@ -238,7 +270,7 @@ class AmazonModel(LLM):
             if self.supports_cache:
                 body["system"].append({"cachePoint": self.cache_control})
 
-        if self.reasoning:
+        if self.reasoning and self.max_tokens:
             if self.max_tokens < 1024:
                 self.max_tokens = 2048
             budget_tokens = kwargs.pop(
@@ -251,9 +283,10 @@ class AmazonModel(LLM):
                 }
             }
 
-        inference: dict[str, Any] = {
-            "maxTokens": self.max_tokens,
-        }
+        inference: dict[str, Any] = {}
+
+        if self.max_tokens:
+            inference["maxTokens"] = self.max_tokens
 
         # Only set temperature for models where supports_temperature is True.
         # For example, "thinking" models don't support temperature: https://docs.claude.com/en/docs/build-with-claude/extended-thinking#feature-compatibility