Lemonade is a local AI server aimed at providing optimized large language model services on users' own GPUs and NPUs, enabling local operation to achieve the same capabilities as cloud APIs, while emphasizing being free and private.
Main forms Lemonade offers two ways to use it. Lemonade Server installs a service that can connect a large number of applications through standard OpenAI, Anthropic, and Ollama APIs. Embeddable Lemonade is a portable binary that can be packaged into your own application, providing it with multimodal local AI and automatically optimizing for the user's PC.
Main capabilities Supports chat, coding, speech, and image generation. Lemonade supports multiple LLMs (GGUF, FLM, and ONNX), whisper, stable diffusion, and other models, and can run on CPU, GPU, and NPU. The built-in Model Manager can be used to browse and download models, and models can also be downloaded using lemonade pull. Custom GGUF/ONNX models can be pulled from Hugging Face or ModelScope, and their source is retained for subsequent updates. Supports setting aliases for models to achieve environment-independent naming and primary/backup failover. For hybrid deployments, an experimental feature allows Lemonade to route requests to any OpenAI-compatible cloud provider (such as Fireworks, OpenAI, OpenRouter, Together, etc.) in addition to local models.
How to use Installation methods include Windows, Linux, macOS, Docker, and installation from source. Mobile provides iOS, Android, and source code. You can try chat, image generation, speech generation, and more through the built-in interface. Configure the base URL of an OpenAI-compatible client library to http://localhost:13305/v1 to connect to Lemonade Server. Embeddable Lemonade provides an embedding guide for developers to integrate into their own applications.
Use cases Users who want to run chat, coding, speech, and image generation models locally on their own PC, and let existing applications access local models through standard APIs; and developers who want to package local multimodal AI into their own applications.
Limitations and notes Cloud Offload is labeled as an experimental feature. The backend and hardware requirements of each inference engine differ, and some backends are labeled as experimental or guest backends. Information on pricing, charges, a specific hardware requirements list, performance data, and features specifically for Roblox games is not currently provided. The project uses the Apache 2.0 license, and some content is licensed as described in the LICENSE.
