cfsutil ai: Build, Profile, and Deploy AI Models via CLI

The cfsutil ai command group provides everything you need to integrate machine learning models into Analog Devices embedded targets, without opening the IDE. You can compile TensorFlow Lite models to deployable C/C++ source files, verify operator and memory compatibility before committing to a build, profile per-layer resource usage, and manage models within a workspace .cfsconfig file — all from the command line or a CI pipeline.

Windows dependency: On Windows, all cfsutil ai commands require the Visual C++ Redistributable for Visual Studio 2015 to be installed. Install this redistributable before running any cfsutil ai command.

Available subcommands

Subcommand	Description
`cfsutil ai build`	Compile a model into deployable C/C++ source code
`cfsutil ai compat`	Check model compatibility with a target SoC and core
`cfsutil ai profile`	Profile model resource usage on target hardware
`cfsutil ai model add`	Add a model to a workspace `.cfsconfig`
`cfsutil ai model list`	List models configured in a workspace
`cfsutil ai model update`	Update model properties in a workspace
`cfsutil ai model remove`	Remove a model from a workspace
`cfsutil ai workspace create`	Generate a full workspace from a model file
`cfsutil ai backends list`	List available AI backends (tflm, izer) and their extension fields
`cfsutil ai clean-cache`	Clear the cache of remotely downloaded model and dataset files

Pass --help at any level for usage details:

cfsutil ai --help
cfsutil ai build --help
cfsutil ai model --help

Global options

The following options apply to all cfsutil ai commands.

Flag	Description
`--format text\|json`	Output format. `text` (default) or `json` for machine-readable output

The --json flag was deprecated in CFS 2.2.0. Use --format json instead.

Data model search paths

The ai build, ai compat, and ai profile commands use a .cfsdatamodels index to look up SoC, core, and accelerator data. A default data model is bundled with the installation. To use a custom data model, pass -x / --search-path pointing to a directory containing your generated index file.

cfsutil ai build --search-path /path/to/data-models --soc MAX32690 --core CM4 --model model.tflite

cfsutil ai build

cfsutil ai build [--format text|json] [--config <value>] [-s <value>] [-p <value>] [-c <value>] \
  [-a <value>] [-m <value>] [--dataset <value>] [--cwd <value>] [-o <value>] \
  [--network-config <value>] [-b <value>] [-e <value>...] [-x <value>...] [--ignore-cache]

Compiles an AI/ML model for an embedded target, generating C or C++ source and header files. Supports two build modes: config-file mode (using --config) and flag-driven mode (using --model, --soc, --core).

Build from a `.cfsconfig` file

cfsutil ai build --config .cfs/max78002-csbga.cfsconfig

When --config is provided, --model, --soc, and --core are ignored. This mode builds all models defined in the configuration file at once. Sample output:

Created file "m4/src/adi_tflm/hello_world_f32.cpp"
Created file "m4/src/adi_tflm/hello_world_f32.hpp"
Created model "hello_world_f32"
Created file "m4/src/adi_tflm/adi_tflm.hpp"

Build from flags

Flag	Short	Description
`--soc`	`-s`	Target SoC (for example, `MAX32690`, `MAX78002`, `ADSP-SC835`)
`--core`	`-c`	Target processor core (for example, `CM4`, `FX`, `CM33`)
`--model`	`-m`	Path or URL to the model file (`.tflite`, `.pth.tar`)
`--package`	`-p`	SoC package variant (for example, `WLP`, `TQFN`). Auto-detected if omitted
`--acc`	`-a`	Hardware AI accelerator (for example, `CNN` for MAX78002)
`--output-path`	`-o`	Output directory for generated files, relative to `--cwd`
`--cwd`		Change the working directory (base for relative paths in `.cfsconfig`)
`--dataset`		Path or URL to a sample dataset for quantization calibration
`--backend`	`-b`	Name of the backend to use. Required if `--extension` is provided
`--extension`	`-e`	Backend-specific `key=value` pairs. Can be repeated
`--network-config`		Path or URL to the Izer network configuration YAML. Required for `izer` backend
`--search-path`	`-x`	Additional search path for templates and data models. Can be repeated
`--ignore-cache`		Bypass cache and force a re-download of remote files

# Basic TFLM build
cfsutil ai build --model model.tflite --soc MAX32690 --core CM4

# With explicit package
cfsutil ai build --model model.tflite --soc MAX32690 --package WLP --core CM4

# SHARC FX core
cfsutil ai build --model model.tflite --soc ADSP-SC835 --core FX

# Custom output directory
cfsutil ai build --config .cfs/max78002-csbga.cfsconfig --output-path ./custom_folder

# Run from a different working directory
cfsutil ai build --config /path/to/project/.cfs/max78002-csbga.cfsconfig --cwd /path/to/project

CNN accelerator (izer backend) builds

To target the MAX78002 CNN accelerator, supply --acc CNN and a --network-config YAML file:

cfsutil ai build \
  --soc MAX78002 --core CM4 --acc CNN \
  --model https://raw.githubusercontent.com/analogdevicesinc/ai8x-synthesis/develop/trained/ai85-catsdogs-qat8-q.pth.tar \
  --network-config https://raw.githubusercontent.com/analogdevicesinc/ai8x-synthesis/develop/networks/cats-dogs-hwc-no-fifo.yaml

Pass backend-specific extension flags with -e:

cfsutil ai build \
  --soc MAX78002 --core CM4 --acc CNN \
  --model path/to/model.pth.tar \
  --network-config path/to/config.yaml \
  --backend izer \
  -e Softmax=True \
  -e InputShape=1,32,32

Use cfsutil ai backends list --name izer to see all available extension fields for the izer backend.

JSON output for CI

cfsutil ai build --config .cfs/max32690-tqfn.cfsconfig --format json

Sample JSON output:

[
  {
    "level": "INFO",
    "msg": "Created file \"m4/src/adi_tflm/hello_world_model_f32.cpp\"",
    "event": {
      "status": "OK",
      "type": "FILE",
      "value": "m4/src/adi_tflm/hello_world_model_f32.cpp"
    }
  },
  {
    "level": "INFO",
    "msg": "Created model \"hello_world_model_f32\"",
    "event": {
      "status": "OK",
      "type": "MODEL",
      "value": "hello_world_model_f32"
    }
  }
]

cfsutil ai compat

cfsutil ai compat [--format text|json] [-s <value>] [-p <value>] [-c <value>] \
  [-a <value>] [-m <value>] [-d <value>] [--report-file <value>] [-x <value>...] [--ignore-cache]

Checks whether a machine learning model can run on a supported processor by validating operator support, memory constraints, and data type compatibility. Currently supports TFLM models only (izer/PyTorch models on MAX78002 are not supported).

Flag reference

Flag	Short	Description
`--soc`	`-s`	Required. Target SoC
`--core`	`-c`	Required. Target core
`--model`	`-m`	Required. Path or URL to the model file
`--package`	`-p`	SoC package variant
`--acc`	`-a`	Target accelerator
`--dataset`	`-d`	Path or URL to dataset for compatibility analysis
`--report-file`		Path to write a JSON report file
`--format`		Output format: `text` (default) or `json`
`--search-path`	`-x`	Additional search path for data models. Can be repeated
`--ignore-cache`		Bypass cache and fetch latest remote files

cfsutil ai compat --soc MAX32690 --core CM4 --model model.tflite
cfsutil ai compat --soc MAX32690 --core CM4 --package WLP --model model.tflite
cfsutil ai compat --soc MAX32690 --core CM4 --model model.tflite --format json
cfsutil ai compat --soc MAX32690 --core CM4 --model model.tflite --report-file report.json

The report covers three categories of issues:

Memory constraint issues — flash/RAM overflows (CRITICAL or WARNING severity)
Operator compatibility issues — unsupported operators with suggested alternatives
Data type compatibility issues — unsupported data types (for example, FLOAT32 when INT8 is required)

cfsutil ai profile

cfsutil ai profile [--format text|json] [-s <value>] [-p <value>] [-c <value>] \
  [-a <value>] [-m <value>] [--report-file <value>] [--report-format text|json] \
  [-x <value>...] [--ignore-cache]

Estimates memory usage, hardware performance, per-layer efficiency, and optimization opportunities for a model on the target device. Currently supports TFLM models only.

Flag reference

Flag	Short	Description
`--soc`	`-s`	Required. Target SoC
`--core`	`-c`	Required. Target core
`--model`	`-m`	Required. Path or URL to the model file
`--package`	`-p`	SoC package variant
`--acc`	`-a`	Target accelerator
`--report-file`		Path to write the report file
`--report-format`		Report file format: `json` (default) or `text`
`--format`		Console output format: `text` (default) or `json`
`--search-path`	`-x`	Additional search path for data models. Can be repeated
`--ignore-cache`		Bypass cache and fetch latest remote files

cfsutil ai profile --soc MAX32690 --core CM4 --model model.tflite
cfsutil ai profile --soc MAX32690 --core CM4 --model model.tflite --report-file profile.json --report-format json

The profile report covers five sections: Model Summary, Memory Analysis, Hardware Performance, Per-Layer Performance, and Optimization Opportunities (layerwise memory and MAC suggestions).

cfsutil ai model

Manage AI models within a workspace .cfsconfig file. All model subcommands require either --config <path> or --workspace <path>. When both are provided, --config takes precedence.

ai model add

cfsutil ai model add --core <value> -m <value> [--format text|json] \
  [-w <value>] [--config <value>] [--dataset <value>] \
  [-a <value>] [--network-config <value>] [-e <value>...] [--ignore-cache]

Adds a model entry to the AIModels array in the .cfsconfig file. The model name is derived from the filename (without extension). A model with the same name cannot be added twice.

Flag	Short	Description
`--model`	`-m`	Required. Path or URL to the model file
`--core`		Required. Target core (for example, `CM4`)
`--config`		Path to the `.cfsconfig` file
`--acc`	`-a`	Target accelerator (for example, `CNN`)
`--extension`	`-e`	Backend-specific `key=value` pairs. Can be repeated
`--dataset`		Path or URL to a sample dataset
`--network-config`		Path or URL to Izer network config YAML. Required for izer backends
`--ignore-cache`		Bypass cache and fetch latest remote files
`--format`		Output format: `text` or `json`

# Add a TFLM model
cfsutil ai model add --config .cfs/max32690.cfsconfig --core CM4 \
  --model path/to/model.tflite

# Add with custom symbol and memory section
cfsutil ai model add --config .cfs/max32690.cfsconfig --core CM4 \
  --model path/to/hello_world_f32.tflite \
  -e Symbol=hello_world_model -e Section=.data

# Add an izer model for MAX78002
cfsutil ai model add --config .cfs/max78002.cfsconfig --core CM4 \
  --acc CNN \
  --model path/to/model.pth.tar \
  --network-config path/to/network.yaml \
  -e Softmax=true

ai model list

cfsutil ai model list [--format text|json] [-w <value>] [--config <value>] [-c <value>] [--verbose]

Flag	Short	Description
`--config`		Path to the `.cfsconfig` file
`--core`	`-c`	Filter models by target core
`--verbose`		Show backend, files, and extension details
`--format`		Output format: `text` or `json`

cfsutil ai model list --config .cfs/max32690-tqfn.cfsconfig
cfsutil ai model list --config .cfs/max32690-tqfn.cfsconfig --core CM4 --verbose

ai model update

cfsutil ai model update --name <value> --set <value>... [--format text|json] \
  [-w <value>] [--config <value>] [-c <value>] [--ignore-cache]

Updates properties of an existing model using --set key=value. Can be repeated. Updatable core fields: name, enabled, model tflm extension fields: Symbol, Section, ArenaSize, ArenaSection, Dataset, DatasetSection izer extension fields: Softmax, Timer, Prefix, AvgPoolRounding, ClockDivider, InputShape, Fifo, NetworkConfig

cfsutil ai model update --config .cfs/max32690-tqfn.cfsconfig \
  --name hello_world_f32 --set name=hello_world_renamed

cfsutil ai model update --config .cfs/max78002-csbga.cfsconfig \
  --name cats-dogs --set Softmax=true --set ClockDivider=4

ai model remove

cfsutil ai model remove --name <value> [--format text|json] [-w <value>] [--config <value>]

cfsutil ai model remove --name hello_world_f32 --config .cfs/max32690-tqfn.cfsconfig

cfsutil ai workspace create

cfsutil ai workspace create -o <value> --name <value> --soc <value> --board <value> \
  --core <value> -m <value> [-s <value>...] [--dataset <value>] [-f]

Generates a complete CFS workspace from a model file. It runs a compatibility check, creates the workspace, integrates the model, and performs an initial build — all in one command.

Flag	Short	Description
`--output`	`-o`	Required. Output directory (workspace created inside it)
`--name`		Required. Workspace name
`--soc`		Required. SoC name
`--board`		Required. Board name
`--core`		Required. Target core
`--model`	`-m`	Required. Path or URL to the model file
`--dataset`		Path or URL to a sample dataset
`--search-path`	`-s`	Additional search path for plugins and data models. Can be repeated
`--skip-compat`	`-f`	Skip the compatibility check before workspace creation

cfsutil ai workspace create \
  -o c:/tmp \
  --name myNewWorkspace \
  --soc MAX32690 \
  --board EvKit_V1 \
  --core CM4 \
  --model c:/models/model.tflite

Using --skip-compat bypasses hardware validation. The workspace will be created even if the model is incompatible with the target hardware. Use only when you are certain the model is supported.

cfsutil ai backends list

cfsutil ai backends list [--format text|json] [-n <value>]

Lists all available AI backends, or shows detailed information about a specific backend including supported hardware targets and extension fields.

# List all backends
cfsutil ai backends list

# Inspect the izer backend in detail
cfsutil ai backends list --name izer

# JSON output
cfsutil ai backends list --format json

Sample output (all backends):

tflm:  Generation for Tflite-micro run-time
izer:  ai8x-izer generation for MAX78002 CNN

cfsutil ai clean-cache

Clears the local cache of remotely downloaded model and dataset files. Use this to force re-downloading of updated remote files.

cfsutil ai clean-cache

cfsutil

Documentation Index

​Available subcommands

​Global options

​Data model search paths

​cfsutil ai build

​Build from a .cfsconfig file

​Build from flags

​CNN accelerator (izer backend) builds

​JSON output for CI

​cfsutil ai compat

​Flag reference

​cfsutil ai profile

​Flag reference

​cfsutil ai model

​ai model add

​ai model list

​ai model update

​ai model remove

​cfsutil ai workspace create

​cfsutil ai backends list

​cfsutil ai clean-cache

Build docs developers (and LLMs) love

Available subcommands

Global options

Data model search paths

cfsutil ai build

Build from a `.cfsconfig` file

Build from flags

CNN accelerator (izer backend) builds

JSON output for CI

cfsutil ai compat

Flag reference

cfsutil ai profile

Flag reference

cfsutil ai model

ai model add

ai model list

ai model update

ai model remove

cfsutil ai workspace create

cfsutil ai backends list

cfsutil ai clean-cache