Model Config Lookup

Model Config Lookup

Do Not Use

Do not use this Skill when:

• The user only wants to discuss model performance or compare models qualitatively.
• The user wants to install or deploy a model locally (e.g., Ollama, vLLM).
• The user asks about API pricing only, without needing configuration fields.

Workflow

Step 1: Identify the Model

Extract from the user's request:

• Provider name (e.g., DeepSeek, OpenAI, Anthropic, Google)
• Model name or ID (e.g., deepseek-v4-flash, claude-sonnet-4-20250514, gpt-4.1)
• If the user only gives a vague description (e.g., "最新的 Claude"), ask which specific model.

Step 2: Fetch Official Specs — MUST Visit Official Source

铁律：所有数值必须来自官方网站或官方 API 文档，不得仅凭博客、教程或第三方聚合站填写。

Priority order for data sources:

1. Official API documentation (highest priority)
2. Official developer platform / console
3. Official blog / announcement post
4. Hugging Face model card (if open-sourced)
5. Third-party aggregator (lowest, must note source)

Step 2a: Visit the Official Docs Page

Use WebFetch to directly fetch the provider's official documentation page. Known official doc URLs:

上述 URL 可能随官方站点更新而变更，若 WebFetch 无法获取有效内容，跳转 Step 2c 使用 WebSearch 重新查找。

If the provider is not in the list above, use WebSearch to find their official docs:

{provider} API documentation model list
{provider} 官方 API 文档 模型列表

Step 2b: Extract These Fields from Official Source

Always look for these specific fields:

• Model ID: the exact string used in API calls (e.g., deepseek-v4-flash)
• Context Window: total input context in tokens
• Max Output Tokens: maximum single-response output in tokens
• API Base URL: the provider's OpenAI-compatible endpoint
• Multimodal support: whether the model accepts image/vision input
• Function/Tool Calling: whether supported

Step 2c: Cross-Verify

If the official site does not clearly state a value, check secondary sources in priority order. Do NOT guess — if uncertain, state the uncertainty and provide the best-known value with a source note. When values conflict, prefer the official source and note the discrepancy.

Step 3: Output the Configuration Table

Output a Markdown table with exactly these fields. The user copies this into a form or config file:

Field rules:

• API格式: OpenAI Chat Completions 格式 for most providers (DeepSeek, OpenAI, Google Gemini, Qwen, MiMo, most Chinese providers). Anthropic Messages 格式 for Claude models via Anthropic's native API.
• 自定义请求地址: The base URL. The tool automatically appends /chat/completions to this address, so do NOT include /chat/completions, /v1/chat/completions, or any trailing slash. Common values:
  • DeepSeek: https://api.deepseek.com
  • OpenAI: https://api.openai.com/v1
  • Anthropic (OpenAI-compatible): https://api.anthropic.com/v1 — 注意原生端点是 /messages 而非 /chat/completions，选 Anthropic Messages 格式 以使用正确端点
  • Google Gemini: https://generativelanguage.googleapis.com/v1beta/openai
  • Aliyun Qwen: https://dashscope.aliyuncs.com/compatible-mode/v1
  • Silicon Flow: https://api.siliconflow.cn/v1
  • For other providers, look up their OpenAI-compatible endpoint.
• 模型ID: Use the exact Model ID from the API docs, lowercase with hyphens (e.g., deepseek-v4-flash). This is case-sensitive.
• 是否开启多模态: 是 if the model supports vision/image input (e.g., GPT-4o, Claude Sonnet 4, Gemini 2.5). 否 for text-only models. If uncertain, default to 否 with a note.
• API密钥: Provide a contextual hint based on the user's situation. If the user already has a key configured for this provider, say "同上，共用同一个 Key". Otherwise, provide the URL to the key management page.
• 模型展示名称: Human-readable name for display in the model list (e.g., DeepSeek V4 Flash, Claude Sonnet 4).
• 上下文窗口（输入）: Total input context window in tokens. Use the numeric value only (e.g., 1048576, not "1M tokens").
• 上下文窗口（输出）: Maximum output tokens in a single response. Use the numeric value only (e.g., 384000).
• 工具调用轮次: 根据供应商文档填写。如果文档未标注，留空或标注 "N/A"。

Step 4: Supplementary Notes

After the table, add a brief section if relevant:

• Thinking/reasoning models: If the model supports a thinking mode (e.g., DeepSeek V4, Claude with extended thinking), note how to disable it if needed for simple tasks.
• Cost reminder: If pricing info was found, add a one-line cost summary.
• Known caveats: Any gotchas (e.g., "V4-Flash requires extra_body={"thinking":{"type":"disabled"}} to avoid empty content on some prompts").

Multi-Model Requests

If the user asks for multiple models at once (e.g., "帮我查一下 Flash 和 Pro"), output one table per model, separated by a horizontal rule. Search for all models in parallel if possible.

Error Handling

• Model not found: If the model name does not match any known provider, ask the user to clarify the provider and model name.
• Specs unavailable: If official docs are unreachable, use the most authoritative secondary source (e.g., model cards on Hugging Face, official blog posts). Note the source.
• Conflicting numbers: If multiple sources give different context window sizes, prefer the official API docs > official blog > Hugging Face model card > third-party aggregator. State the source.