miLLM: Mechanistic Interpretability LLM Server

miLLM is a full-featured, OpenAI API–compatible inference server for running open models locally — in the same family as vLLM and Ollama. What sets it apart is what it does while it serves: miLLM can attach pretrained sparse autoencoders (SAEs) to a running model and steer its behavior in real time, turning mechanistic-interpretability features into a live control surface for generation.

What makes it different

• Lock to a model — load and pin a specific model for stable, repeatable serving.
• Attach SAEs to layers — hook one or more pretrained sparse autoencoders onto one or more layers of that model.
• Stage features for steering — line up multiple interpretable features to influence generation during inference.
• Blend influences — when more than one staged feature is set to a non-zero strength, miLLM blends their effects simultaneously, so you can compose behaviors rather than toggle a single knob.

In practice

Because miLLM speaks the OpenAI API, it drops in as the local inference backend for tools you already use — for example Open WebUI. From there you can dynamically shape responses by adjusting the steering strength of selected features, watching the model’s behavior shift live as you turn features up or down.

Part of a workflow

miLLM is the serving half of a two-application mechanistic-interpretability ecosystem. Its companion, miStudio, handles the discovery side — training SAEs, discovering and labeling features, and testing and exporting them. miLLM then operationalizes those features at inference time in a production-style local serving environment. Each application stands on its own, but they are considerably more powerful used together: miStudio finds and proves the features; miLLM puts them to work.

Explore miLLM

• Documentation — miLLM Manual
• Overview video — Steer Open WebUI Chat with miLLM
• Source code — github.com/hitsainet/miLLM