Edgewise model files

ONNX exports that the Edgewise SDK downloads when no official ONNX build exists. Folders are versioned (v1, v2, …) and never overwritten; Edgewise checks each file's SHA-256 after download.

Chronos preprocessing (scaling, patching, masks) runs in TypeScript; see edgewise.json beside each model. Tokenizers are loaded from the original LiquidAI repositories at the same commits. Export scripts: scripts/export/ in the Edgewise repo.

The wasm-q4 folders hold the official 4-bit exports with GatherBlockQuantized rewritten to standard operators, so they run on ONNX Runtime Web's WebAssembly build (which lacks that kernel). Outputs are identical to the originals; configs and tokenizers still come from the original repositories.

d1-omni is split into three graphs: the vision and audio encoders turn media into prefix embeddings, and decide scores each option's <|mask|> marker. Prompt encoding, image tiling and the log-mel front end run in TypeScript. The position-embedding resize weights are passed into vision because ONNX Runtime's WebGPU provider has no antialiased Resize. fp32 exports match PyTorch exactly; 4-bit changed a top answer, so only 8-bit is hosted.

d1-3B is split into an embedding, two decoder halves and a vision graph, so each 8-bit file stays under 2 GB and loads without external data. Its embedding table is full precision: ONNX Runtime quantizes Gather only at 4 bits, which moved answers ten times more. The halves are too large for ONNX Runtime Web's WebAssembly memory, so Edgewise runs d1-3B on Bun and Node only.