model-library 0.1.8__tar.gz → 0.1.9__tar.gz

{model_library-0.1.8 → model_library-0.1.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.8
+Version: 0.1.9
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT

{model_library-0.1.8 → model_library-0.1.9}/model_library/base/base.py

@@ -76,6 +76,7 @@ class LLMConfig(BaseModel):
     top_k: int | None = None
     reasoning: bool = False
     reasoning_effort: str | bool | None = None
+    compute_effort: str | None = None
     supports_images: bool = False
     supports_files: bool = False
     supports_videos: bool = False
@@ -155,6 +156,7 @@ class LLM(ABC):
 
         self.reasoning: bool = config.reasoning
         self.reasoning_effort: str | bool | None = config.reasoning_effort
+        self.compute_effort: str | None = config.compute_effort
 
         self.supports_files: bool = config.supports_files
         self.supports_videos: bool = config.supports_videos

{model_library-0.1.8 → model_library-0.1.9}/model_library/config/alibaba_models.yaml

@@ -37,6 +37,11 @@ qwen-3-max-models:
     supports:
       tools: true
       images: false
+    # only applies for <32K input tokens
+    # TODO: add thresholds
+    costs_per_million_token:
+      input: 1.2
+      output: 6.0
 
   alibaba/qwen3-max-2026-01-23:
     label: Qwen 3 Max Thinking

{model_library-0.1.8 → model_library-0.1.9}/model_library/config/all_models.json

@@ -122,7 +122,10 @@
             "ignored_for_cost": false
         },
         "provider_properties": {},
-        "costs_per_million_token": {},
+        "costs_per_million_token": {
+            "input": 1.2,
+            "output": 6.0
+        },
         "alternative_keys": [],
         "default_parameters": {
             "temperature": 0.7
@@ -259,7 +262,9 @@
                 "write_markup": 1.0
             }
         },
-        "alternative_keys": [],
+        "alternative_keys": [
+            "vercel/zai/glm-4.7"
+        ],
         "default_parameters": {
             "temperature": 1.0,
             "top_p": 1.0
@@ -269,6 +274,94 @@
         "full_key": "zai/glm-4.7",
         "slug": "zai_glm-4.7"
     },
+    "vercel/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": "zai/glm-4.7",
+        "provider_name": "vercel",
+        "full_key": "vercel/zai/glm-4.7",
+        "slug": "vercel_zai_glm-4.7"
+    },
+    "xiaomi/mimo-v2-flash": {
+        "company": "Xiaomi",
+        "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",
+        "open_source": true,
+        "documentation_url": "https://platform.xiaomimimo.com/#/docs/",
+        "properties": {
+            "context_window": 256000,
+            "max_tokens": 64000,
+            "training_cutoff": "December 2024"
+        },
+        "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": 0.1,
+            "output": 0.3,
+            "cache": {
+                "read": 0.01,
+                "write_markup": 1.0
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 0.3,
+            "top_p": 0.95
+        },
+        "provider_endpoint": "mimo-v2-flash",
+        "provider_name": "xiaomi",
+        "full_key": "xiaomi/mimo-v2-flash",
+        "slug": "xiaomi_mimo-v2-flash"
+    },
     "google/gemini-3-flash-preview": {
         "company": "Google",
         "label": "Gemini 3 Flash (12/25)",
@@ -856,7 +949,9 @@
             "available_as_evaluator": false,
             "ignored_for_cost": false
         },
-        "provider_properties": {},
+        "provider_properties": {
+            "supports_compute_effort": true
+        },
         "costs_per_million_token": {
             "input": 5.0,
             "output": 25.0,
@@ -906,7 +1001,9 @@
             "available_as_evaluator": false,
             "ignored_for_cost": false
         },
-        "provider_properties": {},
+        "provider_properties": {
+            "supports_compute_effort": true
+        },
         "costs_per_million_token": {
             "input": 5.0,
             "output": 25.0,
@@ -2210,7 +2307,10 @@
             "ignored_for_cost": false
         },
         "provider_properties": {},
-        "costs_per_million_token": {},
+        "costs_per_million_token": {
+            "input": 1.2,
+            "output": 6.0
+        },
         "alternative_keys": [],
         "default_parameters": {
             "temperature": 0.7
@@ -2244,7 +2344,10 @@
             "ignored_for_cost": false
         },
         "provider_properties": {},
-        "costs_per_million_token": {},
+        "costs_per_million_token": {
+            "input": 1.2,
+            "output": 6.0
+        },
         "alternative_keys": [],
         "default_parameters": {
             "temperature": 0.7
@@ -2611,7 +2714,10 @@
             "ignored_for_cost": false
         },
         "provider_properties": {},
-        "costs_per_million_token": {},
+        "costs_per_million_token": {
+            "input": 1.2,
+            "output": 6.0
+        },
         "alternative_keys": [],
         "default_parameters": {
             "temperature": 0.7
@@ -3920,7 +4026,7 @@
             "tools": true
         },
         "metadata": {
-            "deprecated": false,
+            "deprecated": true,
             "available_for_everyone": true,
             "available_as_evaluator": false,
             "ignored_for_cost": false
@@ -4550,7 +4656,7 @@
             "tools": true
         },
         "metadata": {
-            "deprecated": false,
+            "deprecated": true,
             "available_for_everyone": true,
             "available_as_evaluator": false,
             "ignored_for_cost": false

{model_library-0.1.8 → model_library-0.1.9}/model_library/config/anthropic_models.yaml

@@ -22,7 +22,6 @@ base-config:
   documentation_url: https://docs.claude.com/en/docs/about-claude/models/overview
   default_parameters:
     temperature: 1
-
 claude-4-models:
   base-config:
     supports:
@@ -40,6 +39,8 @@ claude-4-models:
       cache:
         read: 0.5
         write: 6.25
+    provider_properties:
+      supports_compute_effort: true
     alternative_keys:
       - anthropic/claude-opus-4-5-20251101-thinking:
           properties:

{model_library-0.1.8 → model_library-0.1.9}/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-0.1.8 → model_library-0.1.9}/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-0.1.9/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-0.1.8 → model_library-0.1.9}/model_library/config/zai_models.yaml

@@ -51,6 +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"

{model_library-0.1.8 → model_library-0.1.9}/model_library/providers/anthropic.py

@@ -23,6 +23,7 @@ from model_library.base import (
     InputItem,
     LLMBatchMixin,
     LLMConfig,
+    ProviderConfig,
     QueryResult,
     QueryResultCost,
     QueryResultMetadata,
@@ -48,6 +49,11 @@ from model_library.utils import (
 )
 
 
+class AnthropicConfig(ProviderConfig):
+    supports_compute_effort: bool = False
+    supports_auto_thinking: bool = False
+
+
 class AnthropicBatchMixin(LLMBatchMixin):
     """Batch processing support for Anthropic's Message Batches API."""
 
@@ -251,6 +257,8 @@ class AnthropicBatchMixin(LLMBatchMixin):
 
 @register_provider("anthropic")
 class AnthropicModel(LLM):
+    provider_config = AnthropicConfig()
+
     def _get_default_api_key(self) -> str:
         if self.delegate_config:
             return self.delegate_config.api_key.get_secret_value()
@@ -547,13 +555,22 @@ class AnthropicModel(LLM):
         body["max_tokens"] = self.max_tokens
 
         if self.reasoning:
-            budget_tokens = kwargs.pop(
-                "budget_tokens", get_default_budget_tokens(self.max_tokens)
-            )
-            body["thinking"] = {
-                "type": "enabled",
-                "budget_tokens": budget_tokens,
-            }
+            if self.provider_config.supports_auto_thinking:
+                body["thinking"] = {"type": "auto"}
+            else:
+                budget_tokens = kwargs.pop(
+                    "budget_tokens", get_default_budget_tokens(self.max_tokens)
+                )
+                body["thinking"] = {
+                    "type": "enabled",
+                    "budget_tokens": budget_tokens,
+                }
+
+        # effort controls compute allocation for text, tool calls, and thinking. Opus-4.5+
+        # use instead of reasoning_effort with auto_thinking
+        if self.provider_config.supports_compute_effort and self.compute_effort:
+            # default is "high"
+            body["output_config"] = {"effort": self.compute_effort}
 
         # Thinking models don't support temperature: https://docs.claude.com/en/docs/build-with-claude/extended-thinking#feature-compatibility
         if self.supports_temperature and not self.reasoning:
@@ -599,7 +616,17 @@ class AnthropicModel(LLM):
 
         stream_kwargs = {**body}
         if is_anthropic_endpoint:
-            betas = ["files-api-2025-04-14", "interleaved-thinking-2025-05-14"]
+            betas = ["files-api-2025-04-14"]
+            if self.provider_config.supports_auto_thinking:
+                betas.extend(
+                    [
+                        "auto-thinking-2026-01-12",
+                        "effort-2025-11-24",
+                        "max-effort-2026-01-24",
+                    ]
+                )
+            else:
+                betas.extend(["interleaved-thinking-2025-05-14"])
             if "sonnet-4-5" in self.model_name:
                 betas.append("context-1m-2025-08-07")
             stream_kwargs["betas"] = betas

{model_library-0.1.8 → model_library-0.1.9}/model_library/providers/google/google.py

@@ -361,16 +361,21 @@ class GoogleModel(LLM):
         contents: list[Content | None] = []
         finish_reason: FinishReason | None = None
 
+        chunks: list[GenerateContentResponse] = []
+
         async for chunk in stream:
+            chunks.append(chunk)
             candidates = chunk.candidates
             if not candidates:
                 continue
 
             content = candidates[0].content
 
+            meaningful_content = False
             if content and content.parts:
                 for part in content.parts:
                     if part.function_call:
+                        meaningful_content = True
                         if not part.function_call.name:
                             raise Exception(f"Invalid function call: {part}")
 
@@ -387,13 +392,15 @@ class GoogleModel(LLM):
                     if not part.text:
                         continue
                     if part.thought:
+                        meaningful_content = True
                         reasoning += part.text
                     else:
+                        meaningful_content = True
                         text += part.text
 
             if chunk.usage_metadata:
                 metadata = chunk.usage_metadata
-            if content:
+            if content and meaningful_content:
                 contents.append(content)
             if candidates[0].finish_reason:
                 finish_reason = candidates[0].finish_reason
@@ -402,6 +409,7 @@ class GoogleModel(LLM):
             self.logger.error(f"Unexpected finish reason: {finish_reason}")
 
         if not text and not reasoning and not tool_calls:
+            self.logger.error(f"Chunks: {chunks}")
             raise ModelNoOutputError("Model returned empty response")
 
         result = QueryResult(

model_library-0.1.9/model_library/providers/vercel.py

@@ -0,0 +1,34 @@
+from typing import Literal
+
+from pydantic import SecretStr
+
+from model_library import model_library_settings
+from model_library.base import (
+    DelegateConfig,
+    DelegateOnly,
+    LLMConfig,
+)
+from model_library.register_models import register_provider
+
+
+@register_provider("vercel")
+class VercelModel(DelegateOnly):
+    def __init__(
+        self,
+        model_name: str,
+        provider: Literal["vercel"] = "vercel",
+        *,
+        config: LLMConfig | None = None,
+    ):
+        super().__init__(model_name, provider, config=config)
+
+        # https://vercel.com/docs/ai-gateway/sdks-and-apis#quick-start
+        self.init_delegate(
+            config=config,
+            delegate_config=DelegateConfig(
+                base_url="https://ai-gateway.vercel.sh/v1",
+                api_key=SecretStr(model_library_settings.VERCEL_API_KEY),
+            ),
+            use_completions=True,
+            delegate_provider="openai",
+        )

model_library-0.1.9/model_library/providers/xiaomi.py

@@ -0,0 +1,34 @@
+from typing import Literal
+
+from pydantic import SecretStr
+
+from model_library import model_library_settings
+from model_library.base import (
+    DelegateConfig,
+    DelegateOnly,
+    LLMConfig,
+)
+from model_library.register_models import register_provider
+
+
+@register_provider("xiaomi")
+class XiaomiModel(DelegateOnly):
+    def __init__(
+        self,
+        model_name: str,
+        provider: Literal["xiaomi"] = "xiaomi",
+        *,
+        config: LLMConfig | None = None,
+    ):
+        super().__init__(model_name, provider, config=config)
+
+        # https://platform.xiaomimimo.com/#/docs/quick-start/first-api-call
+        self.init_delegate(
+            config=config,
+            delegate_config=DelegateConfig(
+                base_url="https://api.xiaomimimo.com/v1",
+                api_key=SecretStr(model_library_settings.XIAOMI_API_KEY),
+            ),
+            use_completions=True,
+            delegate_provider="openai",
+        )

{model_library-0.1.8 → model_library-0.1.9}/model_library/register_models.py

@@ -170,6 +170,7 @@ class DefaultParameters(BaseModel):
     top_p: float | None = None
     top_k: int | None = None
     reasoning_effort: str | bool | None = None
+    compute_effort: str | bool | None = None
 
 
 class RawModelConfig(BaseModel):
@@ -338,6 +339,10 @@ def _register_models() -> ModelRegistry:
                         copy.slug = key.replace("/", "_")
                         copy.full_key = key
                         copy.alternative_keys = []
+                        copy.provider_properties = ProviderProperties.model_validate(
+                            provider_properties
+                        )
+
                         registry[key] = copy
 
     return registry

{model_library-0.1.8 → model_library-0.1.9}/model_library/registry_utils.py

@@ -1,9 +1,8 @@
+import tiktoken
 from functools import cache
 from pathlib import Path
 from typing import TypedDict
 
-import tiktoken
-
 from model_library.base import (
     LLM,
     LLMConfig,
@@ -235,6 +234,14 @@ def get_model_names(
     )
 
 
+"""
+everything below this comment is included for legacy support of caselaw/corpfin custom models.
+@orestes please remove this as part of the migration to a standard CorpFin harness.
+"""
+
+DEFAULT_CONTEXT_WINDOW = 128_000
+
+
 @cache
 def _get_tiktoken_encoder():
     """Get cached tiktoken encoder for consistent tokenization."""
@@ -257,7 +264,7 @@ def auto_trim_document(
         Trimmed document, or original document if trimming isn't needed
     """
 
-    max_tokens = get_max_document_tokens(model_name)
+    max_tokens = get_max_document_tokens(model_name) or DEFAULT_CONTEXT_WINDOW
 
     encoding = _get_tiktoken_encoder()
     tokens = encoding.encode(document)
@@ -284,5 +291,5 @@ def get_max_document_tokens(model_name: str, output_buffer: int = 10000) -> int:
     # Import here to avoid circular imports
     from model_library.utils import get_context_window_for_model
 
-    context_window = get_context_window_for_model(model_name)
+    context_window = get_context_window_for_model(model_name) or DEFAULT_CONTEXT_WINDOW
     return context_window - output_buffer

{model_library-0.1.8 → model_library-0.1.9}/model_library/utils.py

@@ -1,6 +1,5 @@
 import logging
 from collections.abc import Mapping, Sequence
-
 import httpx
 from anthropic import AsyncAnthropic
 from openai import AsyncOpenAI
@@ -75,16 +74,15 @@ def create_anthropic_client_with_defaults(
     )
 
 
-def get_context_window_for_model(model_name: str, default: int = 128_000) -> int:
+def get_context_window_for_model(model_name: str) -> int | None:
     """
     Get the context window for a model by looking up its configuration from the registry.
 
     Args:
         model_name: The name of the model in the registry (e.g., "openai/gpt-4o-mini-2024-07-18" or "azure/gpt-4o-mini-2024-07-18")
-        default: Default context window to return if model not found or missing context_window
 
     Returns:
-        Context window size in tokens
+        Context window size in tokens (or `None` if not found)
     """
     # import here to avoid circular imports
     from model_library.register_models import get_model_registry
@@ -98,7 +96,6 @@ def get_context_window_for_model(model_name: str, default: int = 128_000) -> int
         return model_config.properties.context_window
     else:
         logger.warning(
-            f"Model {model_name} not found in registry or missing context_window, "
-            f"using default context length of {default}"
+            f"Model {model_name} not found in registry or missing context_window"
         )
-        return default
+        return None

{model_library-0.1.8 → model_library-0.1.9}/model_library.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.8
+Version: 0.1.9
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT

{model_library-0.1.8 → model_library-0.1.9}/model_library.egg-info/SOURCES.txt

@@ -68,6 +68,7 @@ model_library/config/openai_models.yaml
 model_library/config/perplexity_models.yaml
 model_library/config/together_models.yaml
 model_library/config/xai_models.yaml
+model_library/config/xiaomi_models.yaml
 model_library/config/zai_models.yaml
 model_library/providers/__init__.py
 model_library/providers/ai21labs.py
@@ -87,7 +88,9 @@ model_library/providers/openrouter.py
 model_library/providers/perplexity.py
 model_library/providers/together.py
 model_library/providers/vals.py
+model_library/providers/vercel.py
 model_library/providers/xai.py
+model_library/providers/xiaomi.py
 model_library/providers/zai.py
 model_library/providers/google/__init__.py
 model_library/providers/google/batch.py
@@ -129,4 +132,5 @@ tests/unit/test_registry.py
 tests/unit/test_result_metadata.py
 tests/unit/test_retry.py
 tests/unit/test_token_retry.py
-tests/unit/test_tools.py
+tests/unit/test_tools.py
+tests/unit/test_utils.py

model_library-0.1.9/tests/unit/test_utils.py

@@ -0,0 +1,15 @@
+"""Unit tests for model_library/utils.py"""
+
+from model_library.utils import get_context_window_for_model
+
+
+def test_get_context_window_for_existing_model():
+    """Test that context window is correctly fetched for a model that exists."""
+    context_window = get_context_window_for_model("openai/gpt-4o-mini")
+    assert context_window == 128_000
+
+
+def test_get_context_window_for_nonexistent_model():
+    """Test that None is returned for a model that doesn't exist."""
+    context_window = get_context_window_for_model("nonexistent/fake-model-xyz")
+    assert context_window is None

{model_library-0.1.8 → model_library-0.1.9}/uv.lock

@@ -157,7 +157,7 @@ wheels = [
 
 [[package]]
 name = "anthropic"
-version = "0.74.0"
+version = "0.77.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
@@ -169,9 +169,9 @@ dependencies = [
     { name = "sniffio" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/5b/f9/baa1b885c8664b446e6a13003938046901e54ffd70b532bbebd01256e34b/anthropic-0.74.0.tar.gz", hash = "sha256:114ec10cb394b6764e199da06335da4747b019c5629e53add33572f66964ad99", size = 428958, upload-time = "2025-11-18T15:29:47.579Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/eb/85/6cb5da3cf91de2eeea89726316e8c5c8c31e2d61ee7cb1233d7e95512c31/anthropic-0.77.0.tar.gz", hash = "sha256:ce36efeb80cb1e25430a88440dc0f9aa5c87f10d080ab70a1bdfd5c2c5fbedb4", size = 504575, upload-time = "2026-01-29T18:20:41.507Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/61/27/8c404b290ec650e634eacc674df943913722ec21097b0476d68458250c2f/anthropic-0.74.0-py3-none-any.whl", hash = "sha256:df29b8dfcdbd2751fa31177f643d8d8f66c5315fe06bdc42f9139e9f00d181d5", size = 371474, upload-time = "2025-11-18T15:29:45.748Z" },
+    { url = "https://files.pythonhosted.org/packages/ac/27/9df785d3f94df9ac72f43ee9e14b8120b37d992b18f4952774ed46145022/anthropic-0.77.0-py3-none-any.whl", hash = "sha256:65cc83a3c82ce622d5c677d0d7706c77d29dc83958c6b10286e12fda6ffb2651", size = 397867, upload-time = "2026-01-29T18:20:39.481Z" },
 ]
 
 [[package]]