Model Assets

Permanent home for Wavefoundry's independently versioned offline model-set assets.

You need Wavefoundry 1.16.0 or later to use the model set 2 asset. Wavefoundry 1.15.x and
earlier declare model set 1, and setup looks only for the exact asset filename its own declared
set names. An older install therefore does not pick wavefoundry-models-2.zip up at all: dropping
it next to a 1.15.x install is silently ignored rather than mis-installed, so a failing wf setup
keeps failing and the file looks inert. Upgrade to 1.16.0 first, then download.

This release is not a framework release. It carries no wavefoundry-<version>.zip, is never
selected by framework upgrade discovery, and is deliberately not marked "latest". It exists so that
a controlled, air-gapped, or download-blocked environment always has one stable URL for the exact
model set a given Wavefoundry version declares.

Which asset do I need?

Model assets are versioned by model set, not by Wavefoundry release. Each Wavefoundry feature
package declares the exact set it requires, and wf setup accepts only that set.

Model set	Asset	Required by
2	wavefoundry-models-2.zip	Wavefoundry 1.16.0 and later

Wavefoundry 1.15.x and earlier declared model set 1. That set has been retired and is not
distributed here: it contained components that no longer meet Wavefoundry's supplier-lineage
policy. Upgrade to 1.16.0 or later to get a distributed offline model set.

Model set 2 contents

Compatibility fingerprint wf-model-set-2-20260811-arctic-s. Every component is Apache-2.0 and is
redistributed with its upstream attribution and license text inside the archive.

Component	Supplies	Upstream	Revision
embedding-fastembed	model.onnx plus config and tokenizers: the fastembed fallback embedder	Snowflake/snowflake-arctic-embed-s	e596f507467533e48a2e17c007f0e1dacc837b33
embedding-clean-onnx	model_fp16.onnx (GPU) and model_int8.onnx (CPU): the accelerated embedder	Snowflake/snowflake-arctic-embed-s	d3c1d2d433dd0fdc8e9ca01331a5f225639e798f
reranker-clean-onnx	model.onnx, model_fp16.onnx (GPU), model_int8.onnx (CPU): the reranker	Xenova/ms-marco-MiniLM-L-6-v2	a09144355adeed5f58c8ed011d209bf8ee5a1fec

One Arctic S embedder serves both the documents and the code layer; MiniLM L6 remains the reranker.

Why Arctic S appears twice. These are not two models and not a precision split. Wavefoundry
provisions two independent caches from the one repository: the graph fastembed downloads and runs
itself, and Snowflake's clean ONNX export that the accelerated paths use. The accelerated component
carries both precisions, picking FP16 on a supported GPU provider and INT8 on CPU. Both caches ship
because either path may be the one your host takes.

The two revisions are different commits of that same repository, recorded independently by each
cache. The ONNX weights are byte-identical across them, verified against the Hub's published
checksums: the commits differ only in repository metadata.

How to use it

Normal installs need nothing here: wf setup fetches these models from Hugging Face on the first
index build and caches them locally.

If a model download cannot complete, first retry wf setup with network access. If that is not
possible:

1. Download the asset that matches the set your Wavefoundry version declares.
2. Leave it zipped.
3. Put it in your target repository root, ~/, ~/.wavefoundry/, ~/.wavefoundry/dist/, or
   ~/Downloads/.
4. Run wf setup again.

Setup discovers only the exact declared asset and verifies the model set, per-component file hashes,
and licenses before it replaces anything. An archive that fails verification leaves an existing
verified cache untouched.

Full details in the
README.