saksornr/typhoon-ocr2-0.8b-webgpu-q4

Typhoon OCR 2 0.8B WebGPU Q4

Independent browser conversion of typhoon-ai/typhoon-ocr2-0.8b, pinned at revision a21b2da0c61a3c0fba7cdefc5562e8e3cee6e741.

Used by the Typhoon WebGPU Playground. Model files are downloaded on demand, and neural inference runs on the visitor’s WebGPU hardware. Input text, images, and audio stay in the browser.

Model and conversion

Qwen3.5-based Thai/English document parsing using the exact Typhoon OCR 2 0.8B checkpoint. Separate vision, embedding, and cached decoder graphs support browser inference. The playground uses the published full-page parsing prompt and displays raw layout JSON, HTML table markup, and LaTeX as escaped text.

Dense matrices use symmetric block-32 INT4 MatMulNBits; embeddings use GatherBlockQuantized. Activations, normalization, biases, and convolutions remain FP32. Use device: "webgpu" and dtype: "q4"; preserve all graph/external tensor filenames under onnx/.

The exporter reuses the Apache-2.0 ONNX architecture topology of onnx-community/Qwen3.5-0.8B-ONNX, pinned at 8fe9dbe3fb39fe902e18df9be84b6a0ac77ae0f6. Every learned initializer is replaced with exact Typhoon OCR checkpoint values, including folded normalization and recurrent parameters. All 473 source tensors are accounted for; the topology’s original learned weights are not used.

Validation included native FP32 mapping, vision, and cache-decode parity plus hardware-WebGPU inference on a synthetic document. This is not a representative OCR benchmark. Quantization, image resizing, and output token limits can affect extraction accuracy and completeness.

Browser runtime

Transformers.js 4.3.2 with its matching ONNX Runtime Web 1.31.0-dev.20260914-8d85527a0.

Use a current desktop Chrome or Edge with hardware acceleration enabled. Tokenization/preprocessing/decoding can run in JavaScript; neural inference uses WebGPU. The application selects no CPU model or full-precision model fallback. Four-bit refers to the quantized weights; the floating-point exceptions are described above.

All package files are at this repository’s root, with ONNX graphs under onnx/ where applicable. manifest.json records source provenance, quantization, file sizes, and SHA-256 checksums. Exporters and runtime integration are available in the playground’s source repository. This is an independent conversion, not an official Typhoon release, and attribution does not imply endorsement.

License and attribution

Model weights and the reused ONNX architecture topology use Apache License 2.0. See NOTICE for the precise original model/topology revisions, attribution, and modifications.