Written by: Nuno Leiria, Founder & CEO @ Nilo
Key Takeaways
- Lemonade is a local AI inference server that runs models on your hardware and exposes an OpenAI-compatible API at localhost:13305.
- The server supports multiple backends (llama.cpp, whisper.cpp, stablediffusion.cpp, FastFlowLM, Kokoro, vLLM) for text, speech-to-text, image generation, and text-to-speech.
- Hardware needs vary. CPU-only runs anywhere, GPU acceleration uses ROCm or Vulkan, and NPU acceleration needs AMD Ryzen AI 300/400-series chips with XDNA 2.
- Compared with Ollama, Lemonade offers broader multimodal support and NPU acceleration, but you install backends separately and deal with more setup steps.
- Nilo is a browser-native 3D creation platform that skips server setup and lets you generate and export 3D assets directly in your browser.
Generate Your First 3D Character In Nilo With No Download
What Lemonade Server Actually Does On Your Machine
Lemonade Server is a local AI inference server sponsored by AMD and released under the Apache 2.0 license. It runs as a background service on your computer, separate from any browser you use to reach it.
That background service routes your requests to specialized backends, each tuned for a different kind of AI task:
- llama.cpp handles text and chat models in GGUF format, with support for Vulkan, ROCm, CUDA, and Metal hardware. This is usually the first backend you try.
- whisper.cpp handles speech-to-text transcription, so you send audio and get text back.
- stablediffusion.cpp handles image generation, letting you create images from prompts.
- FastFlowLM targets AMD XDNA 2 NPUs (neural processing units, dedicated AI chips on newer AMD processors) for faster LLM inference.
- Kokoro handles text-to-speech output, turning text into spoken audio.
- vLLM supports high-throughput deployments on AMD ROCm hardware when you need more scale.
Lemonade exposes an OpenAI-compatible endpoint at http://localhost:13305/v1, plus Ollama-compatible and Anthropic-compatible layers on the same port. The API key field is required by many client libraries, but the server ignores it unless you set LEMONADE_API_KEY in your environment.
One key naming trap affects a lot of aspiring builders like you. The Lemonade web app is the browser UI served by the local server at /app. That is how you chat with your own models. lemonade.gg is the official website of Lemonade, an AI coding assistant built by Lemonade Labs for Roblox game development. The two tools only share a name.
Skip Backend Installs And Start Creating 3D Assets In Your Browser
Hardware Requirements For Running Lemonade Smoothly
Lemonade supports four execution modes depending on your hardware:
- CPU-only works on any supported processor generation and gives you the baseline path. Model size depends on how much system RAM you have.
- GPU acceleration through llama.cpp covers AMD GPUs via ROCm and other GPUs, including NVIDIA, through Vulkan.
- NPU acceleration is limited to AMD Ryzen AI 300/400-series processors. Those chips include XDNA 2 hardware, which is what the NPU backend actually targets. Older Ryzen AI 7000/8000-series chips cannot use the NPU, although they still run Lemonade on CPU or GPU.
- Hybrid mode splits work between the NPU and integrated GPU, and also requires Ryzen AI 300-series or newer.
InfoWorld notes that Lemonade’s biggest documented omission is NVIDIA-specific GPU support, reaching NVIDIA hardware through Vulkan instead of CUDA, even though Lemonade’s docs list CUDA backend support for some NVIDIA platforms. On hardware where only the Vulkan path applies, performance can lag behind Ollama’s native CUDA path. If hardware compatibility is the part you worry about most, Nilo gives you a simpler route because it runs in your browser with no GPU requirement at all.
Build On Any Device With Nilo Running On WebGPU
How To Connect Lemonade To A Browser Interface
Your browser always acts as the client. Lemonade runs as the engine on your machine. Follow this sequence to get a basic setup working:
- Install Lemonade Server. Use the MSI on Windows, Debian or RPM packages on Linux, or a
.pkgon macOS. - Install a backend with
lemonade backends install llamacpp:cpuorllamacpp:vulkanfor GPU. No inference backends are installed by default, so nothing runs until you add one. - Pull a model with a command such as
lemonade pull Qwen3-0.6B-GGUF. - Confirm the server is running with
lemonade statusorcurl http://localhost:13305/live. - Open the built-in web UI at the default address, or point a third-party interface at the API.
For Open WebUI, run it and create an admin account. Go to Settings → Admin → Connections. Add an OpenAI-compatible connection with the URL http://localhost:13305/api/v1 and any placeholder key, then reload the page. The first message can take tens of seconds while the model loads into memory.
For Firefox, you can point the browser sidebar at the local server, which gives you a documented integration path. If you are already in a browser and want to create instead of configure, Nilo skips the server layer entirely.
Start Creating In Your Browser With Nilo For Free
Lemonade Server vs. Ollama: Which Local Server Fits Your Setup
The table below compares Lemonade and Ollama on license, default port, NPU support, and multimodal features so you can scan for what matters to you.
| Attribute | Lemonade Server | Ollama |
|---|---|---|
| License | Apache 2.0 | MIT |
| Default port | 13305 | 11434 |
| NPU support | Yes, via FastFlowLM on Ryzen AI 300/400 | No |
| Multimodal backends | Text, speech-to-text, text-to-speech, image generation | Text only |
On setup friction, Ollama gives you a simpler out-of-the-box install and a broader built-in model library. Lemonade, by contrast, needs a separate backend install after the server itself, and its GUI exposes relatively few configuration options, so you will probably manage more from the command line.
On API compatibility, both expose OpenAI-compatible endpoints. Lemonade also serves Ollama-compatible and Anthropic-compatible APIs on the same port, so clients written for any of the three can connect directly.
On hardware support, the gap widens. Ollama uses a single llama.cpp backend and does not support NPUs. Lemonade supports AMD XDNA 2 NPUs through FastFlowLM and adds speech-to-text, text-to-speech, and image generation in one install. On NVIDIA hardware, Ollama’s native CUDA path usually wins on performance, while Lemonade reaches NVIDIA through Vulkan.
On licensing, Lemonade Server uses Apache 2.0 and Ollama uses MIT. Lemonade’s NPU engine FastFlowLM is free for non-commercial use only and needs a separate license for commercial deployment.
Skip Local Servers And Create 3D Games Directly In Nilo
When A Local Server Is The Wrong Tool Entirely
Everything above assumes you actually want to run models locally. Many people who search “lemonade ai browser engine” really want a browser-based creation tool instead of a server. If that describes you, a local inference server will slow you down instead of helping.
Nilo is a browser-based 3D creation platform that runs fully in your browser using WebAssembly and WebGPU. You avoid installs, backend configuration, and hardware checks as long as you have a modern browser. You can generate 3D characters, weapons, and props from sketches or prompts, then export to Roblox Studio or other engines without opening a terminal.
The trade-off looks like this:
- Local servers give you privacy, offline operation, and full control. You pay for that control with hardware requirements, backend installation, and port configuration.
- Browser-native engines such as Nilo give you instant access and zero server setup. You only need a modern browser.
They solve different problems, so the right choice depends on what you want to build today.
Jump Into Nilo And Start Building 3D Worlds In Your Browser
How To Decide: A Quick Readiness Checklist
Use this checklist before you commit to a local server or a browser-native engine:
- Do you need AI to run offline or keep data on your machine? If yes, a local server fits that need.
- Do you have compatible hardware? Look for a Ryzen AI 300/400-series chip for NPU acceleration or a supported AMD GPU for ROCm. CPU-only runs everywhere but limits model size.
- Are you ready to manage a server? Backends, ports, model pulls, and firewall rules all become part of your workflow.
- Do you want to create 3D content right now with no setup? If that is your main goal, a browser-native engine gets you there in a single click.
Try Nilo Free And Go From Idea To 3D Scene In Minutes
Common Mistakes And Limitations To Watch For
These are the friction points aspiring builders like you run into most often:
- Confusing Lemonade Server with lemonade.gg. They share a name and nothing else, which leads people to install the wrong tool. One is a local inference server, and the other is a Roblox Studio assistant.
- Assuming the browser is the engine. The browser acts as the client. Every request travels over HTTP to a process running on your machine.
- Installing the server without a backend. Lemonade ships with no inference backends, so nothing runs until you install one.
- Expecting NPU acceleration on older hardware. Ryzen AI 7000/8000-series laptops cannot use the NPU, even though they can still run Lemonade on CPU or GPU.
- Hitting port conflicts. The default port is 13305. If you change it, every client configuration must match the new value.
- First-request latency. Models load lazily, so the first message after a cold start can take a minute while the backend process starts and loads the model into memory.
The safest approach is simple. Check hardware compatibility before you install anything, and see whether a browser-native tool already covers your use case before you commit to running a server.
Avoid Setup Pitfalls And Start Building 3D Worlds In Nilo
Frequently Asked Questions
What Is A Lemonade Server?
As covered above, Lemonade Server is a local inference server that runs on your own hardware. It uses your CPU, GPU, or NPU and exposes an OpenAI-compatible API at the default port. Your browser connects to it as a client instead of acting as the engine itself.
Is Lemonade AI Open-Source?
Yes. Lemonade Server is licensed under Apache 2.0, with source available at github.com/lemonade-sdk/lemonade. Some bundled third-party components use separate licenses listed in the repository’s NOTICE file. The server itself is Apache 2.0, but the NPU backend has a separate commercial license, so check the FastFlowLM licensing note above if you plan to ship products.
Is Lemonade AI Safe?
Lemonade runs locally, so prompts and data stay on your machine by default. The privacy policy states that it will not transfer information to other networked systems unless you request it. Model downloads contact Hugging Face or ModelScope only when you start them. The Lemonade Mobile app makes the same promise: no cloud account, no subscription, no telemetry. Your conversations and server configs stay on your device.
Is Lemonade AI Completely Free?
Yes. There is no paid tier. Lemonade Server is free and open-source under Apache 2.0. The server is free to use, but if you rely on the NPU backend for commercial work, review the FastFlowLM license mentioned earlier.
How Do I Connect Lemonade AI To A Browser Interface?
Start the server and confirm it is running with lemonade status. Then open the built-in web UI at the default port. For Open WebUI, add an OpenAI-compatible connection pointing at http://localhost:13305/api/v1 with any placeholder API key, save, and reload the page. Firefox’s sidebar chatbot can also point at the local server through its integration settings.
What Hardware Does Lemonade AI Need?
CPU-only inference works on any supported processor and gives you the baseline experience. GPU acceleration covers AMD GPUs via ROCm and other GPUs via Vulkan. For NPU acceleration, you need a Ryzen AI 300/400-series chip with XDNA 2, as noted in the hardware section.
Can I Create 3D Games With Lemonade AI?
Lemonade handles asset generation, not world-building. It generates text, images, and speech, and version 11.0.0 added a 3D generation modality by porting the Trellis.2 image-to-3D pipeline and exposing a POST /v1/3d/generations endpoint with a GUI 3D model viewer. It can produce 3D assets, but it will not build or run a full game for you. For browser-based 3D creation, including characters, props, and environments you can export to Roblox or other platforms, you need a dedicated creation engine. Nilo is a browser-native 3D creation engine built for that workflow, with no install required.
Conclusion: Choosing Between A Local Server And A Browser Engine
Lemonade works as a local AI server that you reach through a browser. It needs installation, a compatible backend, and hardware that can handle the load. It suits privacy-focused builders on AMD hardware who want local text, image, and speech generation, especially on Ryzen AI 300/400-series machines where NPU acceleration helps.
If you searched for a “lemonade ai browser engine” because you want to create a world, a character, or a game directly in your browser, a local inference server will not match that goal. Nilo is a browser-native 3D creation platform that runs on WebAssembly and WebGPU with zero install. It is built for aspiring builders and already builders like you who want to go from idea to playable world without touching a terminal. You generate 3D assets from text, sketches, or images. You can rig and animate characters with one click, then export directly to Roblox Studio. All of this happens in your browser on almost any device.
If you need local AI on your own hardware, install Lemonade. If you want to build 3D games in your browser right now, open Nilo and start creating.
Join Nilo’s Open Beta And Try Building And Playing For Free
