# Model Strategy / 模型策略 This document explains the example purchase and routing idea. It is not a benchmark claim, affiliate recommendation, or requirement for using the plugin. 本文说明示例中的购买与路由思路,不是模型排行榜、商业推广,也不是使用插件的前置要求。 ## Principle / 原则 Use a lower-cost general model for most traffic, upgrade only when the task justifies it, and call media models explicitly. Deterministic routing avoids paying another model merely to decide which model should answer. 大多数请求先交给成本较低的通用模型,只有任务确实需要时才升级,媒体模型则明确调用。确定性路由也避免为了“决定让谁回答”再额外调用一次模型。 ```text Ordinary request / 普通请求 ──> DeepSeek V4 Flash Complex reasoning / 复杂推理 ──> DeepSeek V4 Pro Design work / 设计工作 ───────> Kimi K3 Coding work / 编码工作 ───────> Qwen3.7 Plus Image request / 生图请求 ─────> text orchestrator / 文字模型调度 ──> Seedream 5.0 Lite ``` ## Example responsibilities / 示例职责 ### DeepSeek V4 Flash Default route for daily questions, summaries, lightweight analysis, and general orchestration. In the example deployment it uses `high`, but effort support and pricing belong to the provider and must be checked live. 作为日常问答、汇总、轻量分析和一般调度的默认路由。示例部署使用 `high`,但推理强度是否支持及其价格由提供方决定,必须以实时目录为准。 ### DeepSeek V4 Pro Reserved for architecture decisions, difficult reasoning, critical review, and tasks where a wrong answer is more expensive than a higher model call. 用于架构决策、复杂推理、重要审查,以及“回答错误的代价高于模型调用差价”的任务。 ### Kimi K3 Used for product design, frontend direction, visual hierarchy, interaction concepts, content planning, and other design-led work. The plugin adds K3-specific visible-output guidance but never hides returned blocks. 用于产品设计、前端方向、视觉层级、交互概念、内容策划和其他设计主导的工作。插件会增加 K3 专属输出约束,但不会隐藏模型返回的内容块。 ### Qwen3.7 Plus Used for coding implementation, structured tool use, repository tasks, and engineering execution. This is a routing preference, not a claim that it is always the strongest coding model. 用于代码实现、结构化工具调用、仓库任务和工程执行。这只是路由偏好,不代表它在所有情况下都是最强编码模型。 ### Doubao Seedream 5.0 Lite / 豆包 Seedream 5.0 Lite Used only through `image_gen`. The active text model understands the request and calls the tool; Seedream creates the image. The tool result records the real image provider and model. 只通过 `image_gen` 使用。当前文字模型负责理解请求并调用工具,Seedream 负责生成图片;工具结果记录真正的图片提供方和模型。 ## Cost controls / 成本控制 - No-match keeps the current or default model. / 未命中时继续使用当前模型或默认模型。 - Rules upgrade only explicit task categories. / 规则只升级明确的任务类别。 - `model_route` is off unless an allowlist is configured. / 没有配置白名单时不启用 `model_route`。 - Image generation requires an explicit tool call. / 生图必须明确调用工具。 - Acceptance calls are excluded from CI because they may cost money. / 真实模型验收不会进入持续集成,因为可能产生费用。 Do not copy prices from this repository into a purchasing decision. Providers can change prices, quotas, context limits, and model availability. Check the current official pages and start with a small quota. 不要直接根据本仓库中的旧价格做购买决定。提供方可能随时调整价格、额度、上下文限制和模型可用性。请查看当期官网,并从小额度开始。 ## What contributors can improve / 欢迎贡献者改进 - Better rules that avoid broad keyword collisions / 减少宽泛关键词误命中的规则; - measured evaluation cases instead of personal preference / 用可复现评测替代个人偏好; - provider-neutral examples / 更中立的提供方示例; - cost reports that clearly state date, region, and source / 明确日期、地区和来源的成本报告; - graceful behavior when a purchased model is dormant or renamed / 已购模型休眠或改名时的友好处理。