Skip to content

Roadmap

Model Forge prepares models on the operator's own machine: it tracks a pinned upstream revision, removes refusals when the plan says so, quantizes to a VRAM budget, records a quality check, and publishes a model card with checksums. The GUI is for manual use. The CLI speaks JSON so another program can drive it.

Shipped

  • CLI with --json and stable exit codes.
  • Stages: fetch, abliterate, quantize, fit, eval, card, publish, upload.
  • Fit chooses a VRAM/RAM split (GPU, KV in RAM, MoE experts on CPU, encoder offload, sequential load) and records which split each tier assumes. An offload that fits is kept ahead of a smaller quant.
  • Preflight, heartbeats, and forge status.
  • Desktop window for a local run and a remote session.
  • Canvas image tiers (Z-Image) as a separate entry point.
  • Public Hugging Face upload and private Kaggle upload, with sha256 and size checked after the upload. A public upload requires a permissive licence and a complete card.

Next

  • Documentation site generated in CI from the docstrings and from forge --help.
  • Account setup that keeps Hub, Kaggle, and Drive credentials in the OS keyring.
  • Notice when a pinned upstream revision has a newer release.
  • A small client that drives the JSON CLI.

Out of scope

  • Bundling the prepared weights in an end-user app.
  • Telemetry that is on unless the operator turns it on.
  • Abliteration tools whose licence cannot be shipped with this Apache-2.0 tree.