Managing a variety of local LLMs can quickly become a headache when you have to manually select the right model for every specific task. This manual overhead prevents users from building truly autonomous, scalable agentic workflows that require precise model selection. Lemonade 11.5.0 solves this by introducing a sophisticated, automated routing system that steers your requests exactly where they need to go.
The Lemonade Router: Smart Steering for Local LLMs
The core of this release is the complete Lemonade Router. Instead of hardcoding a single model for every prompt, you can now steer requests based on defined policies. This allows you to optimize for speed, reasoning capability, or cost-efficiency automatically. If you are interested in multi-model orchestration, this update brings that level of control to your local environment.
Steering with Advanced Policies
The router supports several distinct policy types, which can be authored via the Lemonade GUI or defined in JSON files. This flexibility allows you to move from simple rule-based logic to complex semantic-similarity or even using a small, fast LLM to act as the decision-maker.
| Policy Type | Best Use Case |
|---|---|
| Rule | Simple, logic-based redirection. |
| Classifier | Categorizing input to trigger specific models. |
| Semantic Similarity | Routing based on the meaning of the user prompt. |
| LLM-as-Router | Using a lightweight LLM to orchestrate larger models. |
Automated Workflows and Job Management
Beyond routing, Lemonade has introduced a server-side job engine. This feature allows clients to post multi-step recipes and manage them through dedicated endpoints. You no longer have to worry about long-running processes failing silently; you can now use the /jobs endpoints to pause, interrupt, resume, or delete jobs as needed.
Expanding Capabilities with MCP
The lemond daemon can now act as an MCP client host. This means it can connect to external stdio MCP servers and discover their tools through admin-gated endpoints. For those looking to maximize the power of MCP servers, this is a massive leap forward for local-first agent development.
Model Discovery and Enhanced Security
Finding the right model just got easier. ModelScope is now searchable alongside Hugging Face within the Model Manager. This includes source-aware model discovery, download, and registration, making it a one-stop shop for your local LLM needs.
Secure Remote Access
For professional environments, security is paramount. The Lemonade CLI now supports HTTPS/TLS connections to secure remote endpoints, honoring the https:// schemes in your LEMONADE_HOST settings. This makes it much easier to manage deployments across a local network securely.
POST /v1/classify
Content-Type: application/json
{
"action": "classify",
"text": "What is the capital of France?"
}Who is Lemonade For?
Whether you are an enthusiast or a professional developer, Lemonade is designed to scale with your needs. You can check the installation guide to get started immediately.
- LLM Enthusiasts: Get up and running with LLMs on your GPU or NPU in minutes.
- Developers: Integrate LLMs into your applications using the Lemonade server and standard APIs.
- Agent Developers: Build complex, multi-step agentic workflows with GAIA and the new router.
Ready to transform your local AI setup? Review the detailed release notes for v11.5.0 and start building more intelligent, automated workflows today!
