QTS browser model manifest

Browser clients cannot rely on a single POSIX directory. Use a versioned manifest describing required blobs, optional URLs, and integrity metadata for caching.

JSON schema (informal)

Top-level object:

Field	Type	Required	Description
schema_version	integer	yes	Must be 1 for this revision.
model_id	string	yes	Human-readable id (e.g. Hugging Face repo name).
revision	string	yes	Opaque version (commit, tag, or content hash prefix).
files	array	yes	List of file descriptors (see below).

Each element of files:

Field	Type	Required	Description
logical_name	string	yes	Stable key used by the loader (main_gguf, vocoder_onnx, reference_codec_onnx, reference_codec_preprocess_json, tokenizer_config, …).
filename	string	yes	Basename on disk / in cache (e.g. qwen3-tts-vocoder.onnx).
sha256	string	no	Hex digest for integrity after download.
size_bytes	integer	no	Expected size; helps resumable fetch UIs.
url	string	no	Direct download URL; omit if bytes come from elsewhere (OPFS upload).

Required logical names for full QTS

Align with native ModelPaths / export docs:

• main_gguf — talker GGUF (any supported quantized name).
• vocoder_onnx — qwen3-tts-vocoder.onnx.
• For ICL voice clone: reference_codec_onnx, reference_codec_preprocess_json (filename qwen3-tts-reference-codec-preprocess.json).

Tokenizer / config.json are typically embedded in GGUF for this stack; if split artifacts are added later, extend logical_name enums in xlai-qts-manifest (shared by xlai_qts_core::browser and xlai-wasm).

Caching recommendations

1. Key cache entries by (model_id, revision, logical_name, sha256).
2. If sha256 is missing, use revision + size_bytes only for best-effort invalidation.
3. Prefer Cache Storage or OPFS for multi-hundred-MB GGUF; keep a small manifest in IndexedDB pointing to OPFS handles.

Rust types

See xlai_qts_manifest / xlai_qts_core::browser for serde-compatible structs and validate_required_files; the xlai-wasm crate (feature qts) depends on xlai-qts-manifest for the same types.