OpenHuman 模型路由配置 — 自动分配推理/快速/视觉模型

OpenHuman 内置三段式模型路由，会根据任务类型自动选择合适的模型。配置得当的话，复杂推理用强模型、日常对话用便宜模型——既省钱又高效。

三档路由机制

默认配置

[models]
# 快速模型——处理日常对话、简单问题
fast = { provider = "openai_compatible", model = "gpt-4o-mini", base_url = "https://api.openai.com/v1", api_key = "sk-xxx" }

# 推理模型——处理复杂任务、代码、分析
reasoning = { provider = "openai_compatible", model = "o1-mini", base_url = "https://api.openai.com/v1", api_key = "sk-xxx" }

# 视觉模型——处理图片和截图
vision = { provider = "openai_compatible", model = "gpt-4o", base_url = "https://api.openai.com/v1", api_key = "sk-xxx" }

省钱配置方案

推荐组合：快速和推理都用 DeepSeek，视觉用 GPT-4o-mini（支持图片输入）：

[models]
fast = { provider = "openai_compatible", model = "deepseek-chat", base_url = "https://api.deepseek.com/v1", api_key = "sk-deepseek" }
reasoning = { provider = "openai_compatible", model = "deepseek-reasoner", base_url = "https://api.deepseek.com/v1", api_key = "sk-deepseek" }
vision = { provider = "openai_compatible", model = "gpt-4o-mini", base_url = "https://api.openai.com/v1", api_key = "sk-openai" }

纯本地方案（Ollama）

[models]
fast = { provider = "ollama", model = "qwen2.5:7b", base_url = "http://localhost:11434" }
reasoning = { provider = "ollama", model = "qwen2.5:7b", base_url = "http://localhost:11434" }
vision = { provider = "ollama", model = "llava", base_url = "http://localhost:11434" }

模型路由如何决定用哪个？

OpenHuman 根据以下因素自动判断：

• 任务类型：编码/分析→推理模型；聊天→快速模型
• 是否有图片附件→视觉模型
• 对话复杂度：连续多轮复杂对话可能升档

自定义规则

你可以在 config.toml 中设置自定义路由规则：

[model_routing]
# 默认使用 fast 模型
default = "fast"

# 当用户提到关键词时用 reasoning
keyword_trigger = ["写代码", "分析", "调试", "架构"]

# 所有视觉任务用 vision
image_task = "vision"

相关阅读

• OpenHuman 接入 DeepSeek / 国产模型
• Ollama 本地模型配置
• TokenJuice 详解