generated: '2026-08-04'
method: searched
source: >-
  https://doc.kneron.com/docs/toolchain/manual_1_overview/,
  https://hub.docker.com/r/kneron/toolchain,
  https://www.kneron.com/developer_center/
description: >-
  Kneron ships no single unified CLI binary. Its first-party command-line surface
  is the model toolchain, distributed as a container image that a developer runs
  interactively, plus a standalone device firmware update tool. Inside the
  container the primary interface is the `ktc` Python package; `kneronnxopt` is
  the one true command-line optimizer. Names below are quoted verbatim from the
  toolchain manual — nothing here is inferred.
name: Kneron Model Toolchain
distribution: container
install:
  - method: docker
    command: docker pull kneron/toolchain:latest
    registry: https://hub.docker.com/r/kneron/toolchain
  - method: docker-run
    command: 'docker run --rm -it -v /mnt/docker:/docker_mount kneron/toolchain:latest'
    note: Mounts a host working directory into the container at /docker_mount.
commands:
  - name: kneronnxopt
    group: model-preparation
    description: >-
      ONNX model optimizer — optimizes a floating-point ONNX model into the form
      the Kneron compiler expects. Also exposed in Python as
      kneronnxopt.optimize().
  - name: ktc
    group: toolchain-api
    description: >-
      The toolchain's Python package (import ktc), not a shell command. Provides
      the compile/inference entry points used across the workflow.
    functions:
      - {name: 'ktc.kneron_inference()', description: Run inference against an ONNX, BIE or NEF model in the simulator.}
      - {name: 'ktc.compile([km])', description: Batch-compile one or more prepared models into the NEF binary format.}
      - {name: 'km.analysis(input_mapping)', description: Quantize a floating-point model into fixed point, producing a BIE file.}
      - {name: 'km.evaluate()', description: Evaluate a prepared floating-point model.}
  - name: w3m
    group: reporting
    description: Terminal browser bundled in the container for viewing the toolchain's HTML analysis reports.
workflow:
  - step: 1
    name: Floating-point model preparation
    detail: Convert to ONNX, optimize with kneronnxopt.optimize(), evaluate with km.evaluate(), sanity-check with ktc.kneron_inference().
  - step: 2
    name: Fixed-point model generation
    detail: Quantize with km.analysis(input_mapping) to produce a BIE file, then validate with ktc.kneron_inference() against the BIE.
  - step: 3
    name: Compilation
    detail: Batch-compile with ktc.compile([km]) to produce a NEF binary, then simulate with ktc.kneron_inference() against the NEF.
platforms: [520, 530, 630, 720, 730]
related_tools:
  - name: Kneron DFUT
    version: 3.2.0
    description: Device firmware update tool for Kneron NPUs; Ubuntu and Windows builds.
    url: https://www.kneron.com/developer_center/
docs: https://doc.kneron.com/docs/toolchain/manual_1_overview/