tigris zoo

Find and download precompiled .tgrs models from the TiGrIS model zoo. Downloads need no account.

Usage

tigris zoo list [OPTIONS]
tigris zoo fetch [MODEL] [OPTIONS]
tigris run downloaded-model/model.tgrs --input downloaded-model/example-input.bin --output prediction.bin
tigris codegen downloaded-model/model.tgrs --format core -o model.c

Filters

list and fetch take the same filters. All are optional; unspecified resources stay unconstrained.

Flag	Description
--model	Model name (fetch takes it as the first argument)
--category	Task category, e.g. time-series-forecasting
--runtime	A runtime release the build must be compatible with, e.g. 0.11.4
--backend	Kernel backend, e.g. reference
--quantization, --quant	Weight and activation precision, e.g. float32
-m, --mem	Maximum fast arena; repeat for the slow arena
-f, --flash	Maximum plan size in bytes

-m and -f bound the plan’s own arenas and size, not total application RAM or flash. The newest published build that matches wins.

tigris zoo list

Flag	Description
--json	Print the matching catalog entries as JSON
-v, --verbose	Show artifact IDs, runtime compatibility and publication details

Without --verbose it prints a table of model, category, precision and memory:

tigris zoo list

model                category               precision    fast RAM    slow RAM         plan
vww-mobilenetv1      Person detection       int8        64.00 KiB   27.03 KiB   283.56 KiB
vww-mobilenetv1      Person detection       int8        10.00 KiB   45.00 KiB   283.97 KiB
resnet8              Image classification   int8        48.00 KiB    3.03 KiB    87.56 KiB
har-cnn1d            Activity recognition   int8         4.00 KiB       800 B    38.00 KiB
kws-ds-cnn           Keyword spotting       int8        16.00 KiB       544 B    39.64 KiB
electricity-hourly   Forecasting            float32      1.00 KiB       768 B    16.19 KiB

With -v, each build gets its artifact ID, runtime range, tested runtime releases, plan schema, backends and publication date:

resnet8-s8-s9-48k-3b7f5d5e956844e0   resnet8, Image classification, int8
  runtime     >= 0.11.1, no known upper bound
  tested      0.11.3, 0.11.4
  schema      9, backends reference
  memory      fast 48.00 KiB, slow 3.03 KiB, plan 87.56 KiB
  published   2026-09-28T06:44:18.327025Z

tigris zoo fetch

Flag	Description
--artifact ID	Fetch an exact build, including a withdrawn one (with a warning)
-o, --output	New destination directory; defaults to the artifact ID

fetch verifies every file hash and never replaces an existing directory.

Global options

Flag	Description
--offline	Use the Hugging Face cache without network requests
--catalog FILE	Read a local zoo snapshot instead of the model repository
--revision	Catalog revision to read; defaults to main

What a download contains

File	Content
model.tgrs	The plan
readme.md	Input and output conventions, normalization, evaluation and limitations
evaluation.json	Task measurements and runtime parity results
example-input.bin, example-output.bin	One input tensor and the reference output the model was validated against
license.txt	The model’s license and data attribution
manifest.json	The build record as published: compiler, sources, file hashes
download.json	The current runtime constraints, tested runtime releases and catalog revision

Use download.json for dependency decisions. Runtime ranges include both ends; a missing maximum means no known upper bound. Tested releases are listed separately, and matching a range does not claim every release in it was tested.

Run a download

On a native wheel, tigris run executes a download directly on your machine with the bundled reference runtime:

tigris zoo fetch electricity-hourly -o downloaded-model
tigris inspect downloaded-model/model.tgrs
tigris run downloaded-model/model.tgrs \
    --input downloaded-model/example-input.bin --output prediction.bin

Before executing, run checks the plan’s checksum against manifest.json and the loaded runtime against the model’s runtime range, taken from download.json (which must belong to the same artifact) or else from the manifest. The model’s readme.md names the reference behind example-output.bin, and evaluation.json records the parity measured against it, exact or within a tolerance, for the runtime releases the model was tested with. Compare prediction.bin with it accordingly.

For a device, the runtime is supplied separately: the C runtime from tigris-runtime or the ESP-IDF component, at a release inside the model’s runtime range.