Public debug instance — higgs-avatar v1.0.0 on Modal

One always-hot instance (min_containers=1, A100) for public checking/debugging. No auth.

Base URL (stable): https://bosonai--avatar.modal.run

Browser check (no tooling)

Player UI: https://bosonai--avatar.modal.run/avatar/video/stream/player — the release's built-in MSE player ("Offline Noninteractive Avatar Stream Player"): pick a ref image + audio in the browser, watch the fMP4 stream live.

curl checks

# is the process up?
curl https://bosonai--avatar.modal.run/health
# replica state (informational — routing is already warmup-gated, no client gate needed)
curl https://bosonai--avatar.modal.run/avatar/video/stream/replica/status
# full route surface (72 paths)
curl https://bosonai--avatar.modal.run/openapi.json

Full API e2e (create → audio → init.fmp4 → fragments → file.mp4)

python3 milestones/v1_e2e_client.py        # prints PASS/FAIL, exit 0/1

Session state semantics

• Routing is warmup-gated at the infrastructure level (2026-07-03): if your request is routed at all, the replica is session-ready. accepts_new_session remains visible for information only.
• init.fmp4 returns 425 until the first segment exists (honors ?wait_ms=), then 200.
• Single-session profile: one render at a time per replica; concurrent clients queue.
• Resolution buckets: 640x640 / 640x480 / 480x640 (640x480 pre-warmed at boot; others pay one-time PCG capture on first use).

Ops

• Deploy config: milestones/wan_higgs_v1.py (image orzeai/higgs-avatar-inference:a100-v1.0.0-5c0cc03, private mirror of the official release tarball).
• Modal dashboard: https://modal.com/apps/bosonai/main/deployed/higgs-avatar-v1
• Redeploy cold start ≈ 12 min (single-bucket warmup); URL never changes (label-pinned).