# Lemonade 2026.45.3: New Desktop UI, Model Routing and Upgrade Checks

Lemonade's stable release redesigns its desktop and web clients, with visual routing controls, request traces and API changes that local-AI users should check.

By Paras Madan | 2026-11-05


Lemonade has released a redesigned desktop app and web interface for its local AI server. The [stable 2026.45.3 release](https://github.com/lemonade-sdk/lemonade/releases/tag/v2026.45.3) was published on November 4 at 16:54 UTC, following an earlier release candidate. It brings model management, visual routing controls and request inspection into a more unified interface.

The release also adds K2-Horizon family and Flint model entries, Chinese localization and a mobile web interface. That last addition makes the browser client usable on smaller screens; it is not an announcement of a native mobile app.

For someone already running local models, the useful change is how much of the surrounding work can be inspected in one place: finding the model, preparing its backend, configuring requests and understanding a failed or slow run.

## A shared desktop and browser interface

The [Lemonade App guide](https://lemonade-server.ai/docs/guide/app/) explains that the desktop and browser clients share a React interface. Navigation is organized around Chat, Models, Backends, Apps, Monitor and Settings, with global search and a download manager.

The division between client and server matters. The server owns model files, loaded models, inference backends and persistent load configuration. Each graphical client keeps its own connection, chat history, layout and preferences. Connecting another browser therefore does not create another independent copy of the model server.

Model availability still depends on the underlying software and hardware. The guide explicitly warns that showing an unsupported backend does not make it compatible. A catalog entry also does not establish that the selected model will fit in a particular computer's available memory.

This is a good reason to check the model's configuration and backend requirements before downloading a large checkpoint, rather than treating the new interface as a promise of universal compatibility.

## Routing becomes easier to inspect

The [interface implementation notes](https://github.com/lemonade-sdk/lemonade/pull/3834) describe a visual editor for collection.router policies. Users can build nested AND, OR and NOT conditions instead of editing the policy entirely as JSON.

That can make a routing rule easier to review with someone who did not write it. It also creates a useful testing question: which model receives a request when several conditions match, or when none of the expected conditions does? A clear diagram should be followed by representative requests and a check of the actual route.

Monitor brings together server health, prompt and generation throughput, time to first token, inference slots, cache use and system resource information. Request traces and server logs help explain what happened during a run. These are tools for investigating performance; the release does not establish a general inference-speed improvement.

## Local-first still requires checking destinations

Lemonade's [cloud offload documentation](https://lemonade-server.ai/docs/guide/configuration/cloud/) describes an experimental option for using cloud providers alongside local models. Those models appear in the server's catalog and can be used by connected clients.

Before using confidential material, check the selected model and provider configuration. A local application window can still send inference to a configured remote service. The relevant privacy question is where the request goes, not just where the interface runs.

The [telemetry guide](https://lemonade-server.ai/docs/guide/telemetry/) adds a separate consideration. Tracing is disabled by default, but users can enable it and export traces to a local collector or a cloud observability service. Traces can contain prompt inputs, model outputs and thinking blocks. The documented input, output and thinking redaction settings default to false.

Choose the collector and redaction settings deliberately before turning tracing on. Local inference and local trace storage are separate configuration decisions.

## Check API clients before upgrading

The HTTP server has also been reorganized, with three documented behavior changes:

- GET requests to specified POST-only paths now return 405 instead of 404.
- Model routes use a consistent missing-model error message.
- Invalid field types in form-encoded requests return 400 instead of 500.

Applications that branch on those status codes or error messages should be tested against the new version. The release provides downloads for Windows, macOS and several Linux distributions, along with embeddable binaries.

A practical upgrade test is small: load a familiar model, run a known request, inspect the result, then exercise the error paths your application handles. Check routing and trace destinations separately. The redesigned interface should make that work easier to follow, while the model, backend and client integration still determine whether the complete setup works well.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI infrastructure and developer-tool updates into clear technical writing and product stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).