logit-loom-models 0.2.1

Pinned optional model profiles and artifact verification for Logit Loom
Documentation

Optional model profiles

Logit Loom does not redistribute model weights. This directory contains an exact, machine-checked acquisition catalog for three maintained experiment profiles:

Profile Intended role Exact download Current integration status
qwen3-0.6b-q8-0 Small text mechanics 609.8 MiB First-class; retained Vulkan acceptance passed
minit2i-b16 Small direct-RGB image mechanics 3.88 GiB First-class; retained Vulkan acceptance passed
krea-2-turbo Advanced latent image mechanics 12.36 GiB First-class; retained Vulkan acceptance passed

“Catalogued” means the repository pins exact upstream commits, file names, byte counts, weight digests, license locations, and local layout. It does not mean that a model-backed acceptance run has passed. That stronger status requires the adapter, opt-in accelerator execution, and captured mechanical receipts described in the model integration plan. The retained Qwen, MiniT2I, and Krea reports passed that execution gate and their catalog entries are first-class. The checker requires every profile that claims passed acceptance to have a matching retained passed report.

Inspect and acquire

The repository-local command uses the current hf CLI. It never accepts a token on the command line; normal Hugging Face authentication, including HF_TOKEN, remains the CLI's responsibility.

cargo run --quiet -p logit-loom-xtask -- models check
cargo run --quiet -p logit-loom-xtask -- models list
cargo run --quiet -p logit-loom-xtask -- models fetch \
  qwen3-0.6b-q8-0 --dir /path/to/model-store --dry-run
cargo run --quiet -p logit-loom-xtask -- models fetch \
  qwen3-0.6b-q8-0 --dir /path/to/model-store
cargo run --quiet -p logit-loom-xtask -- models verify \
  qwen3-0.6b-q8-0 --dir /path/to/model-store
cargo run --quiet -p logit-loom-xtask -- models verify-artifact krea-2-turbo \
  krea-2-turbo-q6-k TURBO/Krea-2-Turbo-Q6_K.gguf \
  --path /path/to/an/existing/Krea-2-Turbo-Q6_K.gguf

--dry-run validates the catalog and prints exact hf download commands. It does not invoke hf or create a directory. A real fetch downloads only the listed files at the pinned 40-character revision, then verifies every byte count and every available SHA-256 digest. An interrupted fetch can be resumed by repeating the same command. The destination must resolve outside the repository checkout so model artifacts cannot accidentally enter a release.

verify-artifact is the no-network path for a file already managed by another local model store. It applies the same byte-count and SHA-256 checks and emits a path-free JSON receipt.

The retained acquisition report in reports/acquisition-2026-07-25.json uses a bounded, versioned Rust contract. It records the hf CLI version, available filesystem bytes, catalog and verified artifact bytes, and exact path-free receipts. All three profiles are fully verified there. Krea uses the mixed acquisition method because its three runtime components were verified across a repository fetch and a caller-managed model store rather than one shared directory.

Artifacts are placed under <destination>/<profile>/<source>/. For example, the text GGUF is:

<destination>/qwen3-0.6b-q8-0/model/Qwen3-0.6B-Q8_0.gguf

Krea 2 Turbo acquisition verifies only the declared executable artifacts:

cargo run --quiet -p logit-loom-xtask -- models fetch krea-2-turbo \
  --dir /path/to/model-store

The advanced profile uses a pinned Q6_K diffusion GGUF, a pinned Q4_K_M Qwen3-VL text encoder, and the pinned Wan 2.1 VAE expected by the maintained stable-diffusion.cpp integration. Upstream legal documents may be retained as inert source content; they are not catalogued artifacts, runtime inputs, or acceptance gates. Quantization changes storage and runtime mechanics; this catalog does not make a quality comparison with the original Diffusers weights.

Trust and execution boundary

  • No fetch runs during tests, CI, packaging, or documentation builds.
  • The catalog permits only exact file paths and commit hashes; wildcard downloads are not used.
  • Weight files require SHA-256 identities. Git-backed configuration files are bound by exact repository revision and byte count.
  • Remote model code is forbidden. In particular, MiniT2I's upstream custom Python pipeline is not downloaded or executed by this tooling. The maintained adapter uses the reviewed, pinned stable-diffusion.cpp companion described in ADR 0001.
  • Model files remain untrusted inputs to native or tensor runtimes. A matching digest establishes artifact identity, not safety, quality, or efficacy.
  • Model and dependency licenses remain upstream terms. The Logit Loom dual license does not relicense them.