Download OpenVINO static models from an online model zoo URL (remote disk HTTP directory listing) using a model table (name/framework/precision) and store artifacts under a local directory.
name
fetch-static-model
description
Download OpenVINO static models from an online model zoo URL (remote disk HTTP directory listing) using a model table (name/framework/precision) and store artifacts under a local directory.
compatibility
Linux shell (bash), network access to the model zoo host, wget, and python3.
Fetch Static Models (from online model zoo)
Use this skill when you have a
base model zoo URL
(HTTP/HTTPS directory) and a
table of models
(name/framework/precision), and you need to download the OpenVINO IR artifacts for specific precisions (e.g., FP16, INT8) into a
local directory
.
Inputs
Online model zoo URL
(base URL)
Example:
https://ov-share-04.iotg.sclab.intel.com/cv_bench_cache/WW09_static_2026.1.0-21155/
Notes:
Prefer a trailing
/
.
The server is expected to allow directory listing (so recursive download works).
Model table
Must contain (at minimum):
model name
model framework
model precision
Accepted table formats:
Simple “colon format” per line (easy to paste):
roberta-base : PT : FP16/INT8
Or a Markdown table:
model name
framework
precision
roberta-base
pytorch
FP16/INT8
Precision parsing rules:
Split by
/
,
,
, or the word
and
(case-insensitive).
Normalize common spellings:
PT
→
pytorch
TF
→
tensorflow
ONNX
→
onnx
Local directory path
(download destination)
Example:
/mnt/xiuchuan/models/debug
The skill stores files
under this directory
.
Tooling
This skill is implemented by a generic script:
.github/fetch_static_model/scripts/fetch_static_models.sh
Agents should prefer calling this script rather than re-implementing wget logic inline.
Expected remote layout (URL template)
Many zoos use an extra framework “variant” folder under
${MODEL}/${FRAMEWORK}/
(e.g.
pytorch/
under
.../pytorch/
, or
tf_frozen/
under
.../tf/
).
This skill’s script auto-detects
${FW_VARIANT}
by reading the directory listing at:
${BASE_URL}/${MODEL}/${FRAMEWORK}/
Then it downloads:
FP16/FP32 OpenVINO IR:
${BASE_URL}/${MODEL}/${FRAMEWORK}/${FW_VARIANT}/${PRECISION}/1/ov/
INT8 OpenVINO IR (common layout):
${BASE_URL}/${MODEL}/${FRAMEWORK}/${FW_VARIANT}/FP16/INT8/1/ov/
Notes:
Some zoos store INT8 IR under
ov/optimized/
(the script downloads the whole
ov/
tree so it still works).
If a specific model/precision path does not exist (404), the script logs a warning and continues.
Download procedure
0) Preconditions
command
-v wget
command
-v python3
mkdir
-p
"
${LOCAL_DIR}
"
Optional (only if your internal TLS uses an untrusted cert): add
--no-check-certificate
to
wget
.
type f ( -name '.xml' -o -name '.bin' ) | head Common issues 403/404 errors: the URL template for that precision may differ on the server. “Too much downloaded”: ensure -np is set and the URL ends at the desired directory. TLS/cert issues: consider --no-check-certificate (internal environments only).