Lemonade v10.8 Released: Turning Local Models into MCP Tools

The Lemonade v10.8 update adds intelligent memory and context management features. It now dynamically manages VRAM to automatically unload idle models and adjust KV-cache size to free up GPU memory, and model pinning lets you keep required models resident in memory.

Key updates include the following:

• Cloud Offload: Supports a provider-agnostic backend that lets you use OpenAI-compatible providers (Fireworks, OpenRouter, Together, etc.) alongside local models.
• MCP Gateway: Exposes local models as MCP (Model Context Protocol) tools, allowing MCP-supporting hosts to call local models as tools instead of cloud APIs.
• LMX-Omni Expansion: Provides controls such as size and step count for image generation, and allows importing custom Omni models from Hugging Face.
• Platform Expansion: Strengthened support for various hardware and OS, including NVIDIA Blackwell (GB10), AMD ROCm (Windows/Linux), and Debian 13 builds.