Local Model Quantization Router

核心功能是根据硬件配置、隐私要求、任务复杂度等条件，推荐本地量化大模型的路由方案、模型家族、量化级别及端点类型，还提供 fallback 建议。典型用于OpenClaw workload运行前的模型选型，适配本地优先或隐私优先部署。可通过CLI参数或JSON输入需求，输出含决策理由的JSON结果，不下载模型也不修改配置。

stevojarvisai-star

热度 30.81

最近更新 2026-05-11 17:46:49

分类：开发依赖：代码执行 / 读写文件应用场景：DevOps

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add local-model-quantization-router

Metadata

name: local-model-quantization-router
description: Recommend local LLM model routes and quantization levels using hardware, privacy, task complexity, context size, and budget constraints. Use for Qwen/Ollama/local endpoint cost-performance decisions.

Local Model Quantization Router

Use this skill to choose between local quantized models and cloud fallback before running OpenClaw workloads.

Workflow

Describe available hardware and task requirements.
Run scripts/local_model_quantization_router.py with CLI flags or JSON input.
Review the recommended model family, quantization level, endpoint, fallback, and risk notes.
Use the output as routing evidence for local-first or privacy-first deployments.

Parameters

--task TEXT: Task summary.
--complexity {simple,standard,complex,critical}.
--privacy {low,normal,high,regulated}.
--vram-gb FLOAT: GPU memory available.
--ram-gb FLOAT: System memory available.
--context-tokens INT: Required context window.
--hardware PATH: Optional JSON with vram_gb, ram_gb, cpu_only.
--output PATH: Optional JSON output path.

Outputs

route: local-only, local-first, hybrid, or cloud-required.
model: Recommended model family.
quantization: Suggested quant level.
endpoint: Suggested local endpoint type.
fallback: Safer fallback when quality or context is insufficient.
reasons: Evidence for the decision.

No model is downloaded and no config file is changed.