whosellm 0.2.7__tar.gz → 0.3.0__tar.gz

{whosellm-0.2.7 → whosellm-0.3.0}/.bumpversion.toml

@@ -2,7 +2,7 @@
 # 文档 / Documentation: https://callowayproject.github.io/bump-my-version/
 
 [tool.bumpversion]
-current_version = "0.2.7"
+current_version = "0.3.0"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

{whosellm-0.2.7 → whosellm-0.3.0}/CLAUDE.md

@@ -112,6 +112,7 @@ LLMeta("gpt-4o-mini")
 - **`Provider`**（`provider.py`）：动态字符串枚举（OpenAI、Anthropic、Zhipu 等），运行时可通过 `add_member()` 扩展。
 - **`ModelFamily`**（`models/base.py`）：动态字符串枚举，运行时可通过 `add_member()` 扩展。
 - **`ModelCapabilities`**（`capabilities.py`）：冻结 dataclass，描述模型功能（thinking、vision、audio、video、function_calling 等）和限制（context_window、max_tokens 等）。
+  - `media_count_limit`（`MediaCountLimit`）：单次请求的媒体数量上限。值域：`0`（确认不支持）、正整数（文档化上限）、`math.inf`（`UNLIMITED`，无文档化上限，仅受 token 预算约束）；键缺失视同不支持（`limit > 0` 判断支持性，`UNLIMITED` 为 truthy）。`count_mode` 区分 `"per_type"`（各类型独立计数）与 `"combined"`（共享总池，如 GLM 系列）。
   - `supports_json_outputs`：模型能否将输出限制为合法 JSON（`response_format: {type: "json_object"}`），但**不约束 JSON 结构**。
   - `supports_structured_outputs`：模型能否按调用方提供的 **JSON Schema** 生成严格符合结构的输出（`response_format: {type: "json_schema", ...}`）。这是更强的能力，蕴含 `supports_json_outputs`。
 - **`ModelFamilyConfig`**（`models/config.py`）：配置 dataclass，定义模式、默认值和特定模型覆盖。创建时自动注册到全局注册表。

{whosellm-0.2.7 → whosellm-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: whosellm
-Version: 0.2.7
+Version: 0.3.0
 Summary: A unified LLM model version and capability management library
 Author-email: JQQ <jqq1716@gmail.com>
 License: MIT

{whosellm-0.2.7 → whosellm-0.3.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "whosellm"
-version = "0.2.7"
+version = "0.3.0"
 description = "A unified LLM model version and capability management library"
 authors = [
     {name = "JQQ", email = "jqq1716@gmail.com"}

{whosellm-0.2.7 → whosellm-0.3.0}/tests/e2e/test_anthropic.py

@@ -233,7 +233,9 @@ CLAUDE_30_MODELS = [
 # 聚合 + 参数化
 # ============================================================================
 
-ALL_MODELS = CLAUDE_51_MODELS + CLAUDE_LATEST_MODELS + CLAUDE_45_MODELS + CLAUDE_41_MODELS + CLAUDE_40_MODELS + CLAUDE_30_MODELS
+ALL_MODELS = (
+    CLAUDE_51_MODELS + CLAUDE_LATEST_MODELS + CLAUDE_45_MODELS + CLAUDE_41_MODELS + CLAUDE_40_MODELS + CLAUDE_30_MODELS
+)
 
 
 @pytest.mark.e2e

whosellm-0.3.0/tests/models/families/test_media_count_limit_config.py

@@ -0,0 +1,74 @@
+# filename: test_media_count_limit_config.py
+"""
+各家族媒体数量限制的集成断言（官方文档来源见配置注释）
+Per-family media count limit integration tests (official docs cited in config comments)
+"""
+
+import math
+
+from whosellm import UNLIMITED
+from whosellm.model_version import LLMeta
+
+OFFICIAL = (
+    # (模型名, 官方图片数量上限/请求, 计数模式)
+    # 来源 / Sources:
+    # Anthropic: platform.claude.com/docs/en/build-with-claude/vision (600/1M ctx, 100/200k ctx)
+    # OpenAI:    developers.openai.com/api/docs/guides/images-vision (1,500)
+    # Gemini:    ai.google.dev/gemini-api/docs/image-understanding (3,600)
+    # Zhipu:     docs.bigmodel.cn（GLM-5.3-Flash 未公开上限）/ zai-org/GLM-skills（GLM-4.6v ≤50）
+    # DeepSeek:  api-docs.deepseek.com/guides/vision (600)
+    # Alibaba:   help.aliyun.com/zh/model-studio/vision (256 URL / 250 base64)
+    ("claude-fable-5-1", 600, "per_type"),
+    ("claude-mythos-5-1", 600, "per_type"),
+    ("claude-opus-5", 600, "per_type"),
+    ("claude-sonnet-5", 600, "per_type"),
+    ("claude-opus-4-8", 600, "per_type"),
+    ("claude-sonnet-4-6", 600, "per_type"),
+    ("gpt-5.6", 1500, "per_type"),
+    ("gpt-5.5", 1500, "per_type"),
+    ("gemini-3.8-flash", 3600, "per_type"),
+    ("gemini-3.7-flash", 3600, "per_type"),
+    ("glm-4.6v", 50, "combined"),
+    ("glm-4.6v-flash", 50, "combined"),
+    ("deepseek-v4-flash-vision-exp", 600, "per_type"),
+    ("qwen3-vl-plus", 256, "per_type"),
+    ("qwen3-vl-flash", 256, "per_type"),
+)
+
+
+class TestLatestModels:
+    """最新/次新模型的官方图片数量上限 / Official image count caps for latest models"""
+
+    def test_official_image_caps(self):
+        for model_name, expected, mode in OFFICIAL:
+            mcl = LLMeta(model_name).capabilities.media_count_limit
+            assert mcl.get("image") == expected, f"{model_name}: {mcl.get('image')} != {expected}"
+            assert mcl.count_mode == mode, f"{model_name}: count_mode mismatch"
+            assert mcl.is_supported("image") is True
+
+    def test_glm_combined_shared_pool(self):
+        """GLM 系列为合并计数：图片/视频等素材共享 50 的总池"""
+        for model_name in ("glm-4.6v", "glm-4.6v-flash"):
+            mcl = LLMeta(model_name).capabilities.media_count_limit
+            assert mcl.count_mode == "combined"
+            assert mcl.combined == 50
+
+    def test_glm_5v_turbo_no_documented_cap(self):
+        """GLM-5.3-Flash（glm-5v-turbo）官方未公开数量上限 → UNLIMITED"""
+        mcl = LLMeta("glm-5v-turbo").capabilities.media_count_limit
+        assert mcl.get("image") is math.inf
+        assert mcl.get("image") is UNLIMITED
+
+
+class TestUnmeasuredModels:
+    """未实测模型：键缺失视同不支持 / Unmeasured models: missing keys treated as unsupported"""
+
+    def test_200k_claude_without_specific_cap(self):
+        """老代模型未采集 → image 为 0（视同不支持），不误报为无上限"""
+        mcl = LLMeta("claude-3-5-haiku").capabilities.media_count_limit
+        assert mcl.get("image") == 0
+        assert bool(mcl.get("image")) is False
+
+    def test_text_only_model(self):
+        mcl = LLMeta("deepseek-v4-flash").capabilities.media_count_limit
+        assert mcl.get("image") == 0

whosellm-0.3.0/tests/test_media_count_limit.py

@@ -0,0 +1,114 @@
+# filename: test_media_count_limit.py
+"""
+MediaCountLimit 单元测试 / MediaCountLimit unit tests
+
+值域约定 / Value domain:
+- 0         确认不支持 / confirmed unsupported
+- 正整数    文档化上限 / documented cap
+- UNLIMITED 无文档化上限，仅受上下文 token 预算约束 / no documented cap (math.inf)
+- 键缺失 = 未实测，视同不支持 / missing key = unmeasured, treated as unsupported
+"""
+
+import math
+from dataclasses import FrozenInstanceError
+
+import pytest
+
+from whosellm import UNLIMITED, MediaCountLimit
+from whosellm.capabilities import ModelCapabilities
+
+
+class TestDefault:
+    """默认行为 / Default behavior"""
+
+    def test_default_per_type_empty(self):
+        mcl = MediaCountLimit()
+        assert mcl.per_type == {}
+        assert mcl.count_mode == "per_type"
+        assert mcl.combined is None
+
+    def test_missing_key_returns_zero(self):
+        """键缺失（未实测）视同不支持 / Missing keys are treated as unsupported"""
+        mcl = MediaCountLimit(per_type={"image": 10})
+        assert mcl.get("video") == 0
+        assert mcl.is_supported("video") is False
+
+    def test_model_capabilities_default(self):
+        caps = ModelCapabilities()
+        assert caps.media_count_limit == MediaCountLimit()
+        assert caps.media_count_limit.get("image") == 0
+
+
+class TestValueDomain:
+    """取值域语义 / Value domain semantics: 0 / int / UNLIMITED"""
+
+    def test_zero_means_unsupported(self):
+        mcl = MediaCountLimit(per_type={"image": 0})
+        assert mcl.get("image") == 0
+        assert mcl.is_supported("image") is False
+        assert bool(mcl.get("image")) is False
+
+    def test_positive_int_is_documented_cap(self):
+        mcl = MediaCountLimit(per_type={"image": 600})
+        assert mcl.get("image") == 600
+        assert mcl.is_supported("image") is True
+        assert bool(mcl.get("image")) is True
+
+    def test_unlimited_is_truthy(self):
+        """UNLIMITED 必须是 truthy——这是 0 与"无上限"在 bool 上下文的唯一区分点"""
+        mcl = MediaCountLimit(per_type={"image": UNLIMITED})
+        assert mcl.get("image") is math.inf
+        assert mcl.is_supported("image") is True
+        assert bool(mcl.get("image")) is True
+
+    def test_bool_never_conflates_zero_and_unlimited(self):
+        assert bool(0) is False
+        assert bool(UNLIMITED) is True
+
+    def test_unlimited_comparison_with_positive_ints(self):
+        assert UNLIMITED > 50
+        assert UNLIMITED > 0
+        assert not UNLIMITED < 600
+
+    def test_is_supported_uses_positive_check(self):
+        """is_supported 基于 limit > 0，区分 0 / 正整数 / UNLIMITED"""
+        assert MediaCountLimit(per_type={"image": 0}).is_supported("image") is False
+        assert MediaCountLimit(per_type={"image": 1}).is_supported("image") is True
+        assert MediaCountLimit(per_type={"image": UNLIMITED}).is_supported("image") is True
+
+
+class TestCountingMode:
+    """计数模式 / Counting mode"""
+
+    def test_per_type_mode_default(self):
+        mcl = MediaCountLimit(per_type={"image": 600})
+        assert mcl.count_mode == "per_type"
+        assert mcl.combined is None
+
+    def test_combined_mode(self):
+        mcl = MediaCountLimit(per_type={"image": 50, "video": 50}, count_mode="combined", combined=50)
+        assert mcl.count_mode == "combined"
+        assert mcl.combined == 50
+        assert mcl.get("image") == 50
+
+
+class TestImmutability:
+    """不可变性 / Immutability (frozen dataclass)
+
+    注意：frozen 为浅冻结——字段引用不可重绑，与项目内既有 list 字段
+    （supported_image_mime_type 等）的语义保持一致。
+    Note: frozen is shallow — field rebinding is rejected, same semantics as
+    existing list fields (supported_image_mime_type etc.).
+    """
+
+    def test_field_reassignment_rejected(self):
+        mcl = MediaCountLimit(per_type={"image": 50})
+        with pytest.raises(FrozenInstanceError):
+            mcl.per_type = {}  # type: ignore[misc]
+
+    def test_default_factory_isolation(self):
+        """default_factory 保证每个实例独立，不共享可变状态"""
+        a = ModelCapabilities()
+        b = ModelCapabilities()
+        assert a.media_count_limit is not b.media_count_limit
+        assert a.media_count_limit.per_type is not b.media_count_limit.per_type

{whosellm-0.2.7 → whosellm-0.3.0}/uv.lock

@@ -968,7 +968,7 @@ wheels = [
 
 [[package]]
 name = "whosellm"
-version = "0.2.7"
+version = "0.3.0"
 source = { editable = "." }
 dependencies = [
     { name = "parse" },

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/__init__.py

@@ -7,15 +7,17 @@
 LLMeta - 统一的大语言模型版本和能力管理库 / A unified LLM model version and capability management library
 """
 
-__version__ = "0.2.7"
+__version__ = "0.3.0"
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import UNLIMITED, MediaCountLimit, ModelCapabilities
 from whosellm.model_version import LLMeta
 from whosellm.models.base import ModelFamily
 from whosellm.provider import Provider
 
 __all__ = [
+    "UNLIMITED",
     "LLMeta",
+    "MediaCountLimit",
     "ModelCapabilities",
     "ModelFamily",
     "Provider",

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/capabilities.py

@@ -7,7 +7,52 @@
 模型能力定义 / Model capability definitions
 """
 
+import math
 from dataclasses import dataclass, field
+from typing import Literal
+
+# 无文档化上限哨兵：该媒体类型官方未给出数量上限，仅受上下文 token 预算约束
+# Sentinel for "no documented limit": the provider documents no count cap,
+# only the context-window token budget constrains the input.
+UNLIMITED = math.inf
+
+MediaCount = int | float  # 正整数 = 文档化上限；float 仅允许 UNLIMITED（math.inf）
+
+
+@dataclass(frozen=True)
+class MediaCountLimit:
+    """
+    单次请求媒体数量上限 / Per-request media count limits
+
+    值域 / Value domain（键缺失 = 未实测，视同不支持）:
+    - 0         确认不支持 / confirmed unsupported
+    - 正整数    文档化上限，单位：个（张/段/...）/ documented cap (count)
+    - UNLIMITED 无文档化上限，仅受上下文 token 预算约束 / no documented cap
+
+    count_mode:
+    - "per_type": 各类型独立计数（如 "image": 600 仅约束图片）/ per-type caps
+    - "combined": 所有媒体共享一个总池（如 GLM 系列图片+视频合并计数）/ shared pool
+    """
+
+    per_type: dict[str, MediaCount] = field(default_factory=dict)
+    count_mode: Literal["per_type", "combined"] = "per_type"
+    # combined 模式的共享上限（per_type 模式下应为 None）
+    # Shared pool cap for "combined" mode (None when count_mode == "per_type")
+    combined: MediaCount | None = None
+
+    def get(self, media_type: str) -> MediaCount:
+        """
+        返回该媒体类型的上限；键缺失（未实测）返回 0（视同不支持）
+        Return the cap for a media type; missing keys (unmeasured) return 0 (treated as unsupported)
+        """
+        return self.per_type.get(media_type, 0)
+
+    def is_supported(self, media_type: str) -> bool:
+        """
+        该媒体类型是否支持（0 > 0 为 False，正整数与 UNLIMITED 均为 True）
+        Whether the media type is supported at all
+        """
+        return self.get(media_type) > 0
 
 
 @dataclass(frozen=True)
@@ -51,6 +96,9 @@ class ModelCapabilities:
     )  # 支持的图片MIME类型 / Supported image MIME types
     supports_image_base64: bool = True  # 是否支持base64编码的图片 / Whether base64-encoded images are supported
 
+    # 单次请求媒体（图片/视频/音频）数量限制 / Per-request media count limits
+    media_count_limit: MediaCountLimit = field(default_factory=MediaCountLimit)
+
     # 视频相关限制 / Video-related limitations
     max_video_size_mb: float | None = None  # 最大视频大小(MB) / Maximum video size in MB
     max_video_duration_seconds: int | None = None  # 最大视频时长(秒) / Maximum video duration in seconds

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/alibaba.py

@@ -7,7 +7,7 @@
 阿里巴巴模型家族配置 / Alibaba model family configurations
 """
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -142,6 +142,9 @@ QWEN = ModelFamilyConfig(
                 supports_structured_outputs=True,
                 max_tokens=32000,
                 context_window=256000,
+                # 官方：公网 URL/本地路径最多 256 张；Base64 250 张（help.aliyun.com/zh/model-studio/vision）
+                # Per docs: 256 images via URL/local path; 250 via base64
+                media_count_limit=MediaCountLimit(per_type={"image": 256}),
             ),
             patterns=[
                 "qwen3-vl-plus-{year:4d}-{month:2d}-{day:2d}",
@@ -213,6 +216,9 @@ QWEN = ModelFamilyConfig(
                 supports_structured_outputs=True,
                 max_tokens=32000,
                 context_window=256000,
+                # 官方：公网 URL/本地路径最多 256 张；Base64 250 张（help.aliyun.com/zh/model-studio/vision）
+                # Per docs: 256 images via URL/local path; 250 via base64
+                media_count_limit=MediaCountLimit(per_type={"image": 256}),
             ),
             patterns=[
                 "qwen3-vl-flash-{year:4d}-{month:2d}-{day:2d}",

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/anthropic.py

@@ -7,7 +7,7 @@
 Anthropic 模型家族配置 / Anthropic model family configurations
 """
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -41,6 +41,9 @@ CLAUDE = ModelFamilyConfig(
         supports_streaming=True,
         max_tokens=64000,
         context_window=200000,
+        # 200k 上下文模型：官方上限 100 张/请求；1M 模型为 600（见各 specific_models）
+        # 200k-context models: 100 images/request per docs; 1M models: 600 (see specific_models)
+        media_count_limit=MediaCountLimit(per_type={"image": 100}),
     ),
     specific_models={
         # Mythos-class（2026-06-09 发布）：全新顶级层级，优先级高于 opus
@@ -59,6 +62,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-fable-5-{snapshot:snapshot}",
@@ -79,6 +85,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-mythos-5-{snapshot:snapshot}",
@@ -103,6 +112,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-fable-5-1-{snapshot:snapshot}",
@@ -125,6 +137,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-mythos-5-1-{snapshot:snapshot}",
@@ -149,6 +164,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-opus-5-{snapshot:snapshot}",
@@ -169,6 +187,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-sonnet-5-{snapshot:snapshot}",
@@ -189,6 +210,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-opus-4-8-{snapshot:snapshot}",
@@ -209,6 +233,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-opus-4-7-{snapshot:snapshot}",
@@ -229,6 +256,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=128000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-opus-4-6-{snapshot:snapshot}",
@@ -249,6 +279,9 @@ CLAUDE = ModelFamilyConfig(
                 supports_computer_use=True,
                 max_tokens=64000,
                 context_window=1000000,
+                # 官方上限：1M 上下文模型最多 600 张图片/请求（https://docs.anthropic.com/en/docs/build-with-claude/vision）
+                # Per docs: up to 600 images per request for 1M-context models
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "claude-sonnet-4-6-{snapshot:snapshot}",

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/deepseek/deepseek_official.py

@@ -16,7 +16,7 @@ Legacy aliases: deepseek-chat / deepseek-reasoner (announced to be deprecated;
 currently map to deepseek-v4-flash non-thinking and thinking modes respectively).
 """
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -93,6 +93,9 @@ DEEPSEEK = ModelFamilyConfig(
                 supports_structured_outputs=False,
                 max_tokens=384_000,
                 context_window=1_000_000,
+                # 官方文档：单请求最多 600 张图片；≥15 张时单图长边降至 4096px（api-docs.deepseek.com/guides/vision）
+                # Per docs: up to 600 images per request; per-image long edge drops to 4096px with 15+ images
+                media_count_limit=MediaCountLimit(per_type={"image": 600}),
             ),
             patterns=[
                 "deepseek-v4-flash-vision-exp",

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/gemini.py

@@ -7,7 +7,7 @@
 Google Gemini 模型家族配置 / Google Gemini model family configurations
 """
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -436,6 +436,9 @@ GEMINI = ModelFamilyConfig(
                 supports_code_interpreter=True,
                 max_tokens=65536,
                 context_window=1048576,
+                # 官方文档：单请求最多 3600 张图片（https://ai.google.dev/gemini-api/docs/image-understanding）
+                # Per docs: up to 3,600 image files per request
+                media_count_limit=MediaCountLimit(per_type={"image": 3600}),
             ),
             patterns=["gemini-3.7-flash"],
         ),
@@ -460,6 +463,9 @@ GEMINI = ModelFamilyConfig(
                 supports_code_interpreter=True,
                 max_tokens=65536,
                 context_window=1048576,
+                # 官方文档：单请求最多 3600 张图片（https://ai.google.dev/gemini-api/docs/image-understanding）
+                # Per docs: up to 3,600 image files per request
+                media_count_limit=MediaCountLimit(per_type={"image": 3600}),
             ),
             patterns=["gemini-3.8-flash"],
         ),

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/openai/openai_gpt_5_5.py

@@ -1,4 +1,4 @@
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -33,6 +33,9 @@ GPT_5_5 = ModelFamilyConfig(
         supports_computer_use=True,
         max_tokens=128_000,
         context_window=1_050_000,
+        # 官方文档：单请求最多 1500 张图片（https://developers.openai.com/api/docs/guides/images-vision）
+        # Per docs: up to 1,500 images per request
+        media_count_limit=MediaCountLimit(per_type={"image": 1500}),
     ),
     specific_models={
         "gpt-5.5": SpecificModelConfig(

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/openai/openai_gpt_5_6.py

@@ -1,4 +1,4 @@
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -40,6 +40,9 @@ GPT_5_6 = ModelFamilyConfig(
         supports_computer_use=True,
         max_tokens=128_000,
         context_window=1_050_000,
+        # 官方文档：单请求最多 1500 张图片（https://developers.openai.com/api/docs/guides/images-vision）
+        # Per docs: up to 1,500 images per request
+        media_count_limit=MediaCountLimit(per_type={"image": 1500}),
     ),
     specific_models={
         "gpt-5.6": SpecificModelConfig(

{whosellm-0.2.7 → whosellm-0.3.0}/whosellm/models/families/zhipu.py

@@ -7,7 +7,7 @@
 智谱 AI 模型家族配置 / Zhipu AI model family configurations
 """
 
-from whosellm.capabilities import ModelCapabilities
+from whosellm.capabilities import UNLIMITED, MediaCountLimit, ModelCapabilities
 from whosellm.models.base import ModelFamily
 from whosellm.models.config import ModelFamilyConfig, SpecificModelConfig
 from whosellm.provider import Provider
@@ -94,6 +94,9 @@ GLM_VISION = ModelFamilyConfig(
                 supports_streaming=True,
                 max_tokens=128000,
                 context_window=200000,
+                # 官方文档未公开单次请求图片数量上限，仅受 1M 上下文 token 预算约束
+                # No documented per-request image cap (only the context-window token budget)
+                media_count_limit=MediaCountLimit(per_type={"image": UNLIMITED}),
             ),
             patterns=[
                 "glm-5v-turbo-{year:4d}-{month:2d}-{day:2d}",
@@ -210,6 +213,9 @@ GLM_VISION = ModelFamilyConfig(
                 supports_streaming=True,
                 max_tokens=128000,
                 context_window=128000,
+                # 官方：单请求最多 50 张图片（zai-org/GLM-skills，图片+视频等素材总量上限）
+                # Per official GLM-skills: up to 50 images per request (combined with other media)
+                media_count_limit=MediaCountLimit(per_type={"image": 50}, count_mode="combined", combined=50),
             ),
         ),
         "glm-4.6v-flash": SpecificModelConfig(
@@ -227,6 +233,9 @@ GLM_VISION = ModelFamilyConfig(
                 supports_streaming=True,
                 max_tokens=128000,
                 context_window=128000,
+                # 官方：单请求最多 50 张图片（zai-org/GLM-skills，图片+视频等素材总量上限）
+                # Per official GLM-skills: up to 50 images per request (combined with other media)
+                media_count_limit=MediaCountLimit(per_type={"image": 50}, count_mode="combined", combined=50),
             ),
             patterns=[
                 "glm-4.6v-flash-{year:4d}-{month:2d}-{day:2d}",