Model manifest

voxprint-model.json is the shared model record agreed with VOXPrint Components. The schema id is voxprint-model/1. It is provisional: Components still has to review the matching MDL change, so Forge does not push the model index unless you pass --push.

This file is not the Plotweaver run export. run-all still writes manifests/<date>-a100.json and manifests/<date>-a100.MODELS.md. Canvas still writes manifests/<date>-canvas-zimage.json and manifests/<date>-canvas-zimage.MODELS.md. Copy those back as before. voxprint-model.json is one document per published model, uploaded to the repo.

Where it is uploaded

The file sits at the repo root, next to README.md, on every mirror that actually receives the publish:

Hugging Face: HF_USER/voxprint-<id> (canvas repos keep their existing slugs).
Kaggle: the private model, in each variation, under the name voxprint-model.json.
Drive: Voxprint/model-forge/<date>/<model>/voxprint-model.json, when Drive is the destination.

The manifest does not list itself under files. index.json stores the sha256 of voxprint-model.json separately.

If Hugging Face and Kaggle fail and Drive is the fallback, the file is rewritten so mirrors contains only the Drive folder. A failed Hub upload is not recorded as a mirror. INT4 lists both Hub and Drive when both are attempted. If Hub succeeds and the later Drive copy fails, the Hub file can still name that Drive folder.

Fields

Every field is required. The SPDX key is license.

Field	Meaning
`schema`	`voxprint-model/1`
`id`	Model id. Hugging Face slugs drop a leading `voxprint-`.
`display_name`	Card title. LLM publishes use `VOXPrint <id>`.
`app_ids`	`plotweaver` and `audiobook-builder` for text models. `canvas` for image, edit, and video.
`task`	`text-generation`, `text-to-image`, `image-to-image`, or `text-to-video`.
`license`	Lowercase SPDX id, such as `apache-2.0`.
`files`	`{path, size, sha256}` for every uploaded file except this manifest.
`mirrors`	`{kind, ref, revision}`. `kind` is `hf`, `kaggle`, `gdrive`, `modelscope`, or `url`.
`tier`	`min_vram_gib`, `recommended_vram_gib`, `gpu_min_cc`.
`runtime`	`engine` is `llama.cpp`, `diffusers`, or `nunchaku`, plus `version_req`.
`base_model`	Upstream repo id. Empty when a dry-run backfill could not read one.
`card_url`	`https://huggingface.co/<repo>` for the mirror that holds this file. An `hf` mirror with `ours` set, or the single `Mitroshenkov87` mirror, wins over an upstream author listed first. One `hf` mirror is that repo. Empty when there is no `hf` mirror.
`flags`	Includes `least-censored` only when the plan sets `refusal_verified = true` after the refusal probe passed. A repo name, `mode = "ablate"` or `already_abliterated` is not enough, and backfill never infers it.
`on_demand`	`true` for these downloadable weights.
`build`	`forge_version`, `tag`, `date`. `tag` is `""` until a release tag is passed in.

files[].path is relative to the repo root and is the same string on every mirror. Nested canvas paths stay nested (checks/grid.avif). The Kaggle variation uploader still flattens / to __ for those nested bytes; the manifest does not rename them. voxprint-model.json itself has no slash, so its name is the same on Hugging Face, Kaggle, and Drive.

Forge does not publish to ModelScope. modelscope and url are in the schema so a later mirror can use them.

VRAM and compute capability

min_vram_gib and recommended_vram_gib are GiB. MDL treats a 24 GB card as about 23.5 GiB. Forge converts a decimal GB budget with gb * 1000 / 1024 and rounds to the nearest 0.5, so 24 GB becomes 23.5 and 16 GB becomes 15.5. The smaller tier is min_vram_gib. The larger tier is recommended_vram_gib. One tier sets both.

gpu_min_cc defaults to 6.0 for llama.cpp and 8.0 for diffusers and nunchaku. INT4 already treats compute capability 8.0 as the floor.

Runtime

Publish	`engine`	`version_req`
LLM GGUF	`llama.cpp`	Plan `llama_cpp_tag` (default `b11476`)
Canvas dtype conversion	`diffusers`	`>=0.37.1,<0.38`
Wan 2.2 GGUF and FP8	`llama.cpp`	`b3962`

The Wan repo also contains FP8 weights. Those files are listed in files. runtime.engine stays llama.cpp because the GGUF quants are the llama.cpp deliverable.

Index

voxprint-forge-cli and forge share this parser.

voxprint-forge-cli index build --from DIR --out index.json
voxprint-forge-cli index build --from DIR --out index.json --revision REPO=40HEX
voxprint-forge-cli index push --index index.json --push
voxprint-forge-cli index host --from DIR --push
voxprint-forge-cli index backfill --out DIR

index build scans DIR for voxprint-model.json and writes:

[{"id": "tiny", "repo": "Mitroshenkov87/voxprint-tiny", "revision": "<40-hex commit of voxprint-model.json>", "sha256": "<sha256 of voxprint-model.json>"}]

repo is the mirror that holds voxprint-model.json: the hf mirror with ours set, or the single mirror under Mitroshenkov87. The first hf mirror may be the upstream author, and that author repo has no descriptor. A manifest with no such mirror is an error. revision is the commit in that repo that contains voxprint-model.json. The build asks Hugging Face for the last commit that touches the file on the published ref, or takes --revision REPO=SHA (repeatable, 40 lowercase hex digits). It does not use the mirror's revision. That field is often a later weights commit, and the descriptor is not in it. If the commit cannot be resolved, the build fails and writes no index. sha256 is the manifest file, not a weight.

index push without --push exits 2 and does not upload. With --push it uploads index.json to the dataset Mitroshenkov87/voxprint-model-index. CI does not pass --push.

index backfill lists Mitroshenkov87/voxprint-* model repos, skips voxprint-model-index, and writes DIR/<owner>/<repo>/voxprint-model.json. It does not upload. A repo with no licence tag, or a file larger than 8 MiB that has no LFS sha256, is skipped and named in the result. VRAM defaults to 16 GB and 24 GB when the file names do not contain 16gb or 24gb.

Official-source models

A model that Forge uses straight from its author (a LoRA base, the captioner, the age-gate judges) has no Mitroshenkov87 repo, and no mirror is made. Its descriptor lists only the author's source: the hf repo at a pinned commit, or a url folder for a file that is not on Hugging Face. Each file has its size and sha256 from that commit. card_url is the author's page.

The descriptor is kept in the index dataset itself:

Mitroshenkov87/voxprint-model-index (dataset)
  descriptors/<id>/voxprint-model.json

The index row keeps the same four keys:

{"id": "z-image", "repo": "datasets/Mitroshenkov87/voxprint-model-index", "revision": "<dataset commit with the descriptor>", "sha256": "<sha256 of voxprint-model.json>"}

repo is the Hub path of the dataset, so the descriptor URL is https://huggingface.co/<repo>/resolve/<revision>/descriptors/<id>/voxprint-model.json. A reader checks that file against sha256 like any other descriptor. The weights still come from the author's URLs listed in the descriptor.

A descriptor is official-source when it has an hf or url mirror, no mirror marked ours, and no mirror under Mitroshenkov87. Publish order:

forge index host --from DIR --push             # one dataset commit; prints "commit"
forge index build --from DIR --out index.json \
  --revision datasets/Mitroshenkov87/voxprint-model-index=<commit>
forge index push --index index.json --push     # signs and uploads, next serial

Without the pin, index build asks the Hub for the last commit that touches descriptors/<id>/voxprint-model.json.

Schema

schemas/voxprint-model.schema.json is JSON Schema draft 2020-12. Tests validate built documents against it. Publish also checks the same required shape in process, so a bad document is not uploaded when the schema package is absent.

App models

manifests/model-forge.models.json (voxprint-forge-models/1) is the per-app list. Required: Z-Image base BF16 (Tongyi-MAI/Z-Image@04cc4abb), Qwen3-VL-4B-Instruct (@ebb281ec), MiVOLO v2 (iitolstykh/mivolo_v2@53393526) and YuNet 2023mar (opencv_zoo @47534e27). Optional: Florence-2-large (@21a599d4), MOSS-Image Base (@1ceede80), Z-Image-Turbo (@f332072a) and DINOv2-base (@f9e44c81). Every entry has a full commit, the file list with size and sha256, and the licence; all are permissive. voxprint_forge.app_models.missing_required() reads this file. The LoRA bases in lora/bases.py use the same commits. scripts/gen_app_models.py regenerates it from the Hugging Face API without a token.