Run SmolVLA locally with OpenEye

    Vision-language-action model that outputs robotic control vectors from visual observations. For autonomous manipulation tasks. OpenEye, the open-source perception CLI, downloads SmolVLA, runs it on your CPU or GPU, and returns structured JSON you can pipe into scripts, agents, and robots.

    TaskVision-language-action
    CLI idsmolvla
    CreatorHuggingFace
    WeightsHuggingFaceTB/SmolVLA-base (~500 MB)
    HardwareCPU, NVIDIA CUDA
    Install extraopeneye-sh[smolvla]

    1. Install

    pipx install "openeye-sh[smolvla]"

    2. Pull SmolVLA

    openeye pull smolvla

    3. Run inference

    Output is JSON on stdout, so it composes with jq and any other Unix tool.

    openeye run smolvla frame.jpg --pretty

    4. Serve it as an API

    Start a local server with REST and WebSocket endpoints for apps, agents, and live camera streams.

    openeye serve smolvla

    See the serve reference, REST endpoints, and Python SDK.

    Other models