Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

release(v4.9.8): H3 draft decoder fetched from a pinned, hash-verified source — the HF repo it came from is gone

mrbizarroe3fb2f2ce9f9
diff --git a/README.md b/README.md
index ebc5cee..4603f24 100644
--- a/README.md
+++ b/README.md
@@ -13,7 +13,7 @@
</p>
-> **Current release: v4.9.7.** Better anonymous fleet telemetry, no new behaviour for you: renders now report which surface queued them (form, Batch, Storyboard, Characters, Image Studio, Retry), a handful of feature-use counts (Storyboard planning, the Editor, CivitAI, the sample character, a training start), whether the update pop-up was answered, and when the queue paused itself — all closed vocabularies, no text, same install id, same Settings toggle to turn it off (see docs/ANALYTICS.md). Also: the stats page shows complete weight-pack downloads instead of a per-file total. Includes v4.9.6 (H3 clips take their sidecars to Trash, #77). Full notes on the [releases page](https://github.com/mrbizarro/phosphene/releases).
+> **Current release: v4.9.8.** One urgent fix: installing or updating Hailuo H3 failed at the "TAE draft decoder" step with "Repository Not Found" — the Hugging Face repo that file came from was deleted. It is now fetched from a pinned commit of the author's GitHub repo and verified by hash (thanks @macstephen on Pinokio). Includes v4.9.7 (better fleet telemetry) and v4.9.6 (H3 clips take their sidecars to Trash). Full notes on the [releases page](https://github.com/mrbizarro/phosphene/releases).
## Overview
diff --git a/VERSION b/VERSION
index 89faebe..de9a551 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-4.9.7
+4.9.8
diff --git a/docs/STATE.md b/docs/STATE.md
index ed851fd..ecabff1 100644
--- a/docs/STATE.md
+++ b/docs/STATE.md
@@ -117,6 +117,19 @@
> update (EditAnything → LTX 2.5 v2.0); `/loras/guide` wrote a real guide
> (6.5 s).
+> **🚨 2026-09-06 — v4.9.8: H3 install/update broken for everyone by a deleted HF repo — fixed.**
+> `madebyollin/taeh3` on Hugging Face is gone (401/404); install_h3.js fetched
+> the 22 MB TAE draft decoder from it, so every H3 install and "Update Hailuo
+> H3" died at that step (Pinokio, @macstephen). New
+> `scripts/pinokio/h3_fetch_tae.py`: pinned GitHub commit of
+> madebyollin/taehv, size + sha256 verified, atomic, idempotent, HF fallback.
+> The LTX live-preview TAE was never at risk (mirrored on weights-ltx25-v1).
+> Lesson: any third-party download in the install path must be mirrored or
+> hash-pinned — HF repos vanish. #62: Piotr's clean-folder run still shows no
+> identity; sidecar shows base = LTX-2.3 Q8 (by design — trainer is 2.3-based
+> and adapters transfer, the sample character proves it); asked him for the
+> adapter file to test here.
+
> **📈 2026-09-05 — v4.9.7: the fleet picture gets its blind spots filled (analytics only).**
> `source` on every render event; `feature_used` (storyboard_plan/export,
> editor_open/export, civitai_download, sample_character, train_start);
diff --git a/install_h3.js b/install_h3.js
index 068111e..37c8fdd 100644
--- a/install_h3.js
+++ b/install_h3.js
@@ -283,7 +283,12 @@ module.exports = {
},
message: [
"mkdir -p '{{cwd}}/mlx_models/hailuo-h3/models/tae'",
- ".venv/bin/python -c \"from huggingface_hub import hf_hub_download; import shutil; p=hf_hub_download('madebyollin/taeh3','taeh3.safetensors'); shutil.copy(p, '{{cwd}}/mlx_models/hailuo-h3/models/tae/taeh3.safetensors'); print('TAE draft decoder ready')\""
+ // The HF repo this used to come from (madebyollin/taeh3) was deleted
+ // in 2026-09 and every H3 install died here with "Repository Not
+ // Found" (Pinokio, @macstephen). The script fetches the same bytes
+ // from a pinned commit of the author's GitHub repo, verifies the
+ // sha256, and keeps HF as a fallback.
+ ".venv/bin/python ../scripts/pinokio/h3_fetch_tae.py '{{cwd}}/mlx_models/hailuo-h3/models/tae/taeh3.safetensors'"
]
}
},
diff --git a/scripts/pinokio/h3_fetch_tae.py b/scripts/pinokio/h3_fetch_tae.py
new file mode 100644
index 0000000..d50abc3
--- /dev/null
+++ b/scripts/pinokio/h3_fetch_tae.py
@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+"""Fetch the H3 draft decoder (madebyollin's TAE for H3) to a target path.
+
+The Hugging Face repo this came from (madebyollin/taeh3) was deleted in
+September 2026 and every H3 install/update died at this step with
+"Repository Not Found" (Pinokio report by @macstephen, who also found the
+file in the author's GitHub repo). The same bytes live at a PINNED commit
+of github.com/madebyollin/taehv (MIT); we fetch from there first, verify
+the sha256 against the copy that shipped for months, and only then place
+it. HF stays as a fallback in case GitHub is unreachable.
+
+Usage: h3_fetch_tae.py <target-path>
+Exit 0 = the file is in place and verified. Non-zero = nothing placed.
+"""
+import hashlib
+import os
+import sys
+import tempfile
+import urllib.request
+
+EXPECTED_SHA256 = "4fd022bfcab08772fe0536b17ea1a3bbb5625be11e397868d1c5d891863d4c13"
+EXPECTED_BYTES = 22709752
+SOURCES = [
+ ("github.com/madebyollin/taehv @ 62f7591 (MIT)",
+ "https://github.com/madebyollin/taehv/raw/62f7591f59dfbb4c3c02b7a621d180a9eeaba26c/safetensors/taeh3.safetensors"),
+ ("huggingface.co/madebyollin/taeh3 (original, may be gone)",
+ "https://huggingface.co/madebyollin/taeh3/resolve/main/taeh3.safetensors"),
+]
+
+
+def main() -> int:
+ if len(sys.argv) != 2:
+ print("usage: h3_fetch_tae.py <target-path>", file=sys.stderr)
+ return 2
+ target = sys.argv[1]
+ if os.path.isfile(target):
+ with open(target, "rb") as f:
+ if hashlib.sha256(f.read()).hexdigest() == EXPECTED_SHA256:
+ print("TAE draft decoder already in place (verified)")
+ return 0
+ print("existing TAE file does not match the known hash - refetching")
+ os.makedirs(os.path.dirname(target) or ".", exist_ok=True)
+ for label, url in SOURCES:
+ print(f"fetching TAE draft decoder from {label} ...")
+ try:
+ with urllib.request.urlopen(url, timeout=120) as r:
+ data = r.read()
+ except Exception as exc: # noqa: BLE001
+ print(f" failed: {exc}")
+ continue
+ digest = hashlib.sha256(data).hexdigest()
+ if len(data) != EXPECTED_BYTES or digest != EXPECTED_SHA256:
+ print(f" rejected: {len(data)} bytes, sha256 {digest[:16]}... (expected "
+ f"{EXPECTED_BYTES} bytes, {EXPECTED_SHA256[:16]}...)")
+ continue
+ fd, tmp = tempfile.mkstemp(dir=os.path.dirname(target) or ".", suffix=".part")
+ with os.fdopen(fd, "wb") as f:
+ f.write(data)
+ os.replace(tmp, target)
+ print("TAE draft decoder ready (verified sha256)")
+ return 0
+ print("ERROR: could not fetch a verified TAE draft decoder from any source; "
+ "drafts will fall back to the full decoder. Re-run 'Install Hailuo H3' "
+ "later.", file=sys.stderr)
+ return 1
+
+
+if __name__ == "__main__":
+ sys.exit(main())