Lemonade 11.0: AMD Local AI Server Adds TTS and 3D

On July 15, 2026, the Lemonade 11.0 release turned AMD's local AI server from an LLM host into a full multimodal creation stack. The Apache 2.0 project, sponsored by AMD and maintained by a community team, now serves text, image generation, text-to-speech, and 3D models from a single OpenAI-compatible endpoint that runs entirely on your own hardware. For creators who want cloud-grade capabilities without cloud bills or data leaving the machine, this is the first free local server to cover all four modalities at once.

The headline additions are a text-to-speech modality built on the OpenMOSS backend, with voice cloning and voice design, and a brand-new 3D generation pipeline ported from Trellis.2. Both ship with dedicated panels in the desktop GUI and REST endpoints, so you can drive them from a script or from a browser tab.

What Shipped in Lemonade 11.0

Lemonade started life as a way to run optimized language models on AMD Ryzen CPUs, Radeon GPUs, and Ryzen AI NPUs. Version 11.0, covered in detail by Phoronix, pushes it well past chat. The release notes list four capability areas that matter for anyone building creative workflows:

Text-to-speech: a new TTS modality using the OpenMOSS backend, including voice-cloning and voice-design models, exposed through a dedicated GUI panel.
3D generation: a new modality that ports the Trellis.2 image-to-3D pipeline, with a POST /v1/3d/generations endpoint and an in-app model viewer.
Wider model sourcing: ModelScope joins Hugging Face as a second remote registry across the CLI, API, and desktop app.
Better hardware handling: MI300X support for ROCm and vLLM, improved APU memory handling, native VRAM detection on Windows, and automatic FastFlowLM NPU installation on Linux.

Everything sits behind the same server process. You point any OpenAI-compatible client at the local /v1 base URL, and route requests to a chat model, an image model, a voice, or the 3D pipeline depending on the endpoint you call.

Lemonade 11.0 desktop GUI showing model panels — Lemonade 11.0 unifies chat, image, speech, and 3D behind one local server.

How Lemonade 11.0 Compares to Other Local Servers

Most popular local runtimes stop at text, and a few add image generation. Lemonade 11.0 is currently alone in covering speech and 3D in the same free package while still targeting AMD NPUs. The table below sizes it against the common alternatives.

Server	LLM chat	Image gen	Text-to-speech	3D generation	Ryzen AI NPU	OpenAI API	License
Lemonade 11.0	Yes	Yes	Yes	Yes	Yes	Yes	Apache 2.0
Ollama	Yes	Limited	No	No	No	Yes	MIT
LM Studio	Yes	No	No	No	No	Yes	Free, closed
Jan	Yes	No	No	No	No	Yes	Open source

The comparison is not just about feature count. Because Lemonade exposes the OpenAI schema, any tool already written against that schema keeps working when you swap the base URL. Backends include llama.cpp, ONNX Runtime GenAI, ROCm, and experimental vLLM, with Vulkan, Metal, and CUDA paths for non-AMD machines.

The Four Modalities, One Endpoint

The reason the multimodal angle matters is workflow chaining. In 11.0 you can generate an image, feed it into the 3D pipeline, and narrate the result without leaving the server. The 3D modality is a direct port of the Trellis image-to-3D research pipeline, exposed as a clean REST call, and the TTS side uses the OpenMOSS speech models for cloning and design. Pulling models from either Hugging Face or ModelScope means you are not locked to a single registry when a weight is only mirrored in one place.

Diagram of image to 3D to voice workflow running locally — Chain image, 3D, and speech steps locally without sending data to a cloud API.

Why AMD Hardware Owners Should Care

The NPU story is the differentiator. On Linux, the FastFlowLM NPU backend now auto-installs, so Ryzen AI machines light up their neural accelerator without a manual driver dance. On Windows, native VRAM detection and CUDA support for Stable Diffusion smooth out the image path, and the MI300X additions extend the server toward workstation and server-class ROCm setups. If you own a recent Ryzen AI laptop, Lemonade is one of the few servers that will actually use the NPU rather than falling back to CPU. The project homepage at lemonade-server.ai tracks which backends map to which chips.

How to Try Lemonade 11.0

Getting a multimodal local stack running takes only a few steps:

Install the server. Grab the installer from the project site or install the lemonade-sdk package, then start the server process. It listens on a local OpenAI-compatible /v1 endpoint.
Pull a chat model. Use the CLI or desktop app to fetch a model from Hugging Face or ModelScope. Point any OpenAI client library at the local base URL to confirm chat works.
Open the TTS panel. Load an OpenMOSS voice, paste text, and generate speech. Clone a reference voice or design a new one from the same panel.
Call the 3D endpoint. Send an image to POST /v1/3d/generations and preview the mesh in the built-in viewer before exporting it.
Chain the steps. Script an image-to-3D-to-narration pipeline that runs fully offline, using the same auth-free local endpoint for every stage.

Ryzen AI laptop running a local AI server on its NPU — On supported Ryzen AI hardware, Lemonade drives the NPU instead of falling back to CPU.

What This Enables for Creators

For a solo creator or a small studio, Lemonade 11.0 collapses four separate tools into one server you control. A product mockup can become a rotatable 3D asset, a script can become narrated audio in a cloned voice, and an idea can become a generated image, all on a single machine with no per-request billing and no uploads. Because it speaks the OpenAI schema, you can drop it into existing agent frameworks, ComfyUI-style pipelines, or your own scripts as a private drop-in for hosted APIs. The privacy win is real for client work where footage, likenesses, or unreleased designs cannot leave the building.

Frequently Asked Questions

Is Lemonade free and open source?

Yes. Lemonade is released under the Apache 2.0 license and is free to use. It is sponsored by AMD but maintained by a community team that includes AMD employees and outside contributors.

Do I need an AMD machine to run it?

No. AMD Ryzen AI hardware gets the best NPU acceleration through FastFlowLM, but Lemonade also ships CUDA, Vulkan, and Metal backends, so it runs on NVIDIA GPUs, Apple Silicon, and generic CPUs. The NPU auto-install is the AMD-specific perk.

What are the new text-to-speech capabilities?

Version 11.0 adds a TTS modality built on the OpenMOSS backend, including voice-cloning and voice-design models, with a dedicated panel in the desktop GUI for generating and shaping speech.

How does the 3D generation work?

Lemonade ports the Trellis.2 image-to-3D pipeline and exposes it as a REST endpoint at POST /v1/3d/generations, with an in-app model viewer so you can inspect the mesh before exporting.

Can I use my existing OpenAI-based tools with it?

Yes. Lemonade serves an OpenAI-compatible API at a local /v1 base URL. Any client or framework already written for the OpenAI schema works after you change the base URL to the local server.

Which operating systems are supported?

Windows 11, macOS, and major Linux distributions including Arch, Debian, Fedora, and Ubuntu, plus a Docker image that now runs unprivileged by default.