← Runners

LemonadeLemonade

AMD / Lemonade 社区AMD / Lemonade community

开源Open source AMDAMD NPUNPU 本地推理Local inference OpenAI兼容OpenAI-compatible partial lifecycle: active 核验Checked 2026-11-03

Lemonade 是 AMD 生态的本地模型服务器：在本机跑模型，提供 OpenAI 兼容 API，支持 llama.cpp 后端，Ryzen AI MAX+ 上另有实验性 vLLM ROCm 后端。

跨大类对照：OpenHands。

An AMD-ecosystem local model server: run models on your machine, expose an OpenAI-compatible API, support the llama.cpp backend and an experimental vLLM ROCm backend on Ryzen AI MAX+.

适合与不适合When to use it, when not to

适合：AMD Ryzen AI / NPU 用户、想要本地 OpenAI 兼容端点、离线或隐私场景。不适合：NVIDIA 独占优化、想要开箱 GUI 聊天。

Fits AMD Ryzen AI / NPU users who want a local OpenAI-compatible endpoint for offline or private work. Does not fit NVIDIA-only optimisation or an out-of-the-box GUI chat app (see LM Studio and similar).

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-11-03

出品属地Origin: 国外出品foreign-made　AMD 生态 / Lemonade 社区（美国）。AMD ecosystem / Lemonade community (US).
国内使用Direct access: 大陆可直连direct OK　lemonade-server.ai 大陆直连未核验；GitHub/安装包可经镜像获取。Direct reachability from mainland not verified; GitHub and installers are available through mirrors.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-11-03 检索）；开源工具未面向大陆公众提供服务，备案不适用。Not in the CAC public registry (searched 2026-11-03); the open-source tool does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 本地模型服务器：从 Hugging Face 拉取模型并在本机运行；支持 llama.cpp 后端， AMD Ryzen AI MAX+ 上另有实验性 vLLM ROCm 后端（官方 FAQ）。
A local model server: pulls models from Hugging Face and runs them on the machine; supports the llama.cpp backend, plus an experimental vLLM ROCm backend on AMD Ryzen AI MAX+ (official FAQ).
运行位置Where it runs: 本地常驻服务：Windows 安装器或完整 SDK；提供 OpenAI 兼容 API，也可与 Ryzen AI 工具链配合。
A long-running local service: Windows installer or the full SDK; exposes an OpenAI-compatible API and can work with the Ryzen AI toolchain.
本地文件Local files: 模型与数据在本地；不依赖云端账号，适合离线/隐私场景，模型缓存目录与更新由本地服务管理。
Models and data stay local; no cloud account dependency, which suits offline and privacy-sensitive use.
关机后的任务Tasks after shutdown: 常驻服务形态，可被其他应用调用；没有「关机后续跑」的官方声明，长任务的重试与恢复由调用方负责。
A long-running service that other apps can call; no official statement of continuing after shutdown.
工具与扩展Tools & extensions: 以推理服务为主：暴露 OpenAI 兼容端点，适合被其他应用调用；工具调用与 MCP 未逐条核验，复杂编排需要外层框架（如 agent SDK 或工作流平台）。
Primarily an inference service; tool-calling and MCP not verified item by item.
上下文与记忆Context & memory: 长上下文按模型与显存/内存配置；具体限制以官方 FAQ 为准，本站未逐条测量各模型的上下文上限， KV cache 由 llama.cpp 或实验性 vLLM 后端决定。
Long context depends on model and VRAM/RAM configuration; limits follow the official FAQ.
权限与限制Permissions & limits: 本地进程无账号体系；权限由宿主环境与调用方决定，多用户隔离需要自己加一层服务包装与鉴权。
A local process with no account system; permissions come from the host environment.
适合什么任务Best for: 适合：AMD Ryzen AI / NPU 用户、想要本地 OpenAI 兼容端点、离线或隐私场景。不适合：NVIDIA 独占优化、想要开箱 GUI 聊天（看 LM Studio 等）。
Fits AMD Ryzen AI / NPU users who want a local OpenAI-compatible endpoint for offline or private work. Does not fit NVIDIA-only optimisation or an out-of-the-box GUI chat app (see LM Studio and similar).

头号误解Top misconception

把 Lemonade 当成 Windows 专属——官方也有 SDK 与 Linux 后端，但硬件支持有矩阵Treating Lemonade as Windows-only — the SDK and Linux backends exist, but hardware support has a matrix
以为所有模型都能 NPU 加速——NPU 支持按型号与模型而定Assuming every model gets NPU acceleration — NPU support depends on model and chip
把实验性 vLLM ROCm 当稳定后端——官方自己标注 experimentalTreating the experimental vLLM ROCm backend as stable — the official announcement marks it experimental

未知项清单Known unknowns

各 Ryzen 型号的 NPU/GPU 加速矩阵
工具调用与 MCP 支持状态
实验性 vLLM ROCm 后端的稳定边界

NPU/GPU acceleration matrix across Ryzen models
Tool-calling and MCP support status
Stability boundaries of the experimental vLLM ROCm backend

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-11-03，内容更新日content updated on 2026-11-03。

类型Type	名称Name	链接Link
docs	Lemonade · 官方 FAQ（后端与硬件支持；2026-11-03 复核官方页）Lemonade · official FAQ (backend and hardware support)	https://lemonade-server.ai/docs/guide/faq/
docs	Lemonade · vLLM ROCm 公告（实验性后端）Lemonade · vLLM ROCm announcement (experimental backend)	https://lemonade-server.ai/news/vllm-rocm.html
docs	AMD Ryzen AI · Lemonade Server 文档AMD Ryzen AI · Lemonade Server docs	https://ryzenai.docs.amd.com/en/latest/llm/server_interface.html
changelog	Lemonade · GitHub Releases（官方 changelog）Lemonade · GitHub Releases (official changelog)	https://github.com/lemonade-sdk/lemonade/releases

实测记录Measured runs

本档未跑吞吐与显存实测；结论只覆盖运行时能力与许可。

This site has not completed throughput or VRAM tests; the conclusions cover runtime capability and licensing only.