Skip to content

feat(surfaces): expose H3 safely across Mold #841

Description

@jamesbrink

Part of the MiniMax H3 native audio-video milestone (tracking: #814). Expose H3 through one capability/request authority across catalog, CLI/server, Studio, desktop, iPhone remote, web, TUI, and Discord—with ordinary model behavior while preserving exact routing and artifact authority.

Catalog and installation

  • Add two task-aware rows/variants for FL2VA and Ref2VA while deduplicating shared Qwen, processor, visual VAE, and AudioVAE components.
  • Built-in manifests outrank ambiguous live Hugging Face repository synthesis; never bulk-download the roughly 498 GB repository containing duplicate original/converted layouts.
  • Show full/pruned/quantized tier, installed/shared component ownership, required disk/host RAM/VRAM, CUDA/attention requirements, source, format, license, and task limits only from supplied metadata.
  • Apply the reviewed narrow acquisition authority: only the two pinned compact manifests are discoverable/downloadable upstream-direct; arbitrary IDs cannot bypass it, and an installed model remains independently subject to host execution capability.
  • Use planModelInstall for every merged fleet row and preserve exact-host download/recovery ownership.

Shared Studio capability UX

  • Add H3 to shared generation capabilities, resolution/duration/frame-offset logic, prompt requirements, conditioning-aware ExpandTask, model filtering, and video/audio budget helpers.
  • Present Output as video with synchronized audio, fixed 24 fps, aligned actual duration, no CFG/negative prompt, and only the sampler/step choices H3 supports.
  • Provide accessible endpoint controls for first-only, last-only, and first+last FL2VA.
  • Provide an ordered mixed-reference editor with 44pt controls, drag/reorder alternatives, kind badges, video-soundtrack association, duration/count budget, structured errors, and no denoise-strength implication.
  • Preserve references/endpoints in templates, prepared expansion, stale-work recovery, exact-host retry, queue rows, gallery provenance, and “use as prompt/source.”

Surface matrix

  • Server and CLI: full authoritative request/placement/errors/progress/output.
  • Desktop and web: full T2VA/FL2VA/Ref2VA authoring and gallery playback/export.
  • iPhone: remote-only parity, Keychain-only host key, bounded native media upload, safe-area/44pt/16px invariants, and authenticated playback/save.
  • TUI and Discord: capability-driven fields, bounded attachment/reference syntax, clear unsupported/limit errors, and synchronized video output.
  • Machines/host detail: actual H3 component inventory, disk usage, download state, queue plan, attention/backend readiness, host RAM, and cancellation.

Acceptance

  • Contract tests prove every client hides/rejects unsupported fields from server capabilities rather than family-string guesses.
  • End-to-end tests preserve ordered references and exact host/task partition through missing-model pull, cancellation, reconnect, retry, and saved metadata.
  • No surface logs/prompts/URLs expose durable API keys or reference media; media tickets remain exact-path and time-limited.
  • Mobile and desktop use shared model-kind labels and show the textual 18+ NSFW badge when catalog metadata supplies it.
  • All frontend, desktop, iOS, web architecture/dead-code, Rust, docs, and release-sync gates pass.

Dependencies

Metadata

Metadata

Assignees

No one assigned

    Labels

    cliCLI (mold-cli) relateddiscordDiscord bot (mold-discord) relatedfeatureNew capability or modeh3MiniMax H3 model familyserverServer (mold-server) relatedtuiTUI (mold-tui) related

    Type

    No type

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions