FLE Studio: A Local Chat and Lab Interface for llama.cpp and FastLocalExperts

FLE Studio is a new local chat and laboratory application designed to streamline the use of llama.cpp and the FastLocalExperts (FLE) engine. The tool addresses the complexity of configuring launch flags by providing a unified interface for running large MoE models with expert caching.

Key Features

• Live Inference Data: The interface displays real-time metrics during generation, including prompt and generation speed, draft/MTP acceptance rates, NVIDIA GPU utilization, and FLE’s expert-cache hit rate.
• Model Management: A built-in loader lists FLE quantizations of Qwen3.8-Flash-Next and supports any GGUF model from Hugging Face. It verifies file compatibility with local hardware before download, supports resume, and checks SHA-256 hashes.
• Launch Profiles: Users can view and save specific command-line configurations. The app automatically selects the appropriate backend (FLE CUDA, llama.cpp CUDA, Metal, Vulkan, or CPU) or connects to an existing llama-server instance.
• Experimentation: An Experiments tab allows for quick, repeatable speed tests to compare different configuration settings side-by-side.

Technical Details and Availability

The application is MIT licensed and requires no external Python installation, as the installer brings its own Python via uv and the app relies only on the standard library. It is currently at version 0.1.0, with the developer noting potential rough edges. Initial testing was conducted on a Linux machine with an RTX PRO 4500, and feedback is particularly sought from Windows users with NVIDIA GPUs.

Future updates on the roadmap include "Studio Forge", a feature enabling users to build custom FLE-style quantizations (using imatrix, per-tensor types, and optional REAP pruning) directly within the application.