Mobilint Model Zoo

SkillDocs & knowledge

Work on the Mobilint Model Zoo package, its legacy compatibility facade, CLI integration, Transformers integrations, MeloTTS, and repository documentation. Use for Model Zoo changes; implement Vision features in the standalone mblt-vision-python repository instead.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Mobilint Model Zoo skill

What this skill tells your AI

The instructions your AI receives, as published by mobilint/mblt-model-zoo in .agents/skills/mblt-model-zoo/SKILL.md and read by ahel’s review.

  1. Read AGENTS.md, run git status --short, and inspect the relevant parser, exports, tests, and pyproject.toml before editing.
  2. Keep mblt_model_zoo.vision and its Vision compilation exports as thin compatibility layers. Compatibility modules must forward to mblt_vision; never copy Vision implementation, model YAMLs, dataset YAMLs, evaluation, or benchmarks into Model Zoo. Limit tests/vision to generic opt-in facade smoke tests; keep implementation-specific Vision tests in mblt-vision-python.
  3. Make new Vision CLI behavior in mblt-vision-python first. The Model Zoo predict, val, and compile handlers must delegate to mblt_vision.cli.
  4. Preserve Model Zoo CLI help and README examples when its CLI integration changes. Pass board-specific target_device through to the standalone packages; do not reintroduce legacy product/artifact selection in Model Zoo. Supported boards are aries-rb, regulus-ra, regulus-rb, regulus-ra-usb, and regulus-rb-usb; the shared runtime floor is mblt-npu-python>=0.1.0 / mobilint-qb-runtime>=1.4.0. Route every *target_device setter (top-level, encoder_, decoder_, base_, draft_, fc_, and the Qwen3-ASR facade) through _rebuild_backend_for_target_device so cross-board overrides switch the destination backend class atomically.
  5. Keep TPS output driven by mblt_model_zoo/cli/tps_table.py. Preserve local conventions in hf_transformers and MeloTTS. Qwen3-VL 8B (regular and Batch16) is unsupported until the shipped MXQ matches the current runtime's expected output ordering.
  6. Start with focused tests. Report unavailable hardware, downloads, and optional extras instead of weakening validation. For docs, run git diff --check.
  7. When package ownership, public API, CLI bridges, or runtime dependencies change significantly, update AGENTS.md, this canonical skill, the Claude entry point when its workflow changes, and relevant documentation in the same change.
  8. Before a release, build from a clean tree and inspect the wheel. Reject legacy Vision implementation paths such as vision/models/, vision/utils/preprocess/, vision/utils/postprocess/, vision/utils/datasets/, and vision/utils/evaluation/.

Signals

GitHub stars
24
Forks
1
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
mblt-model-zoo
Source
github.com/mobilint/mblt-model-zoo