WebGPU Chat graphs

ONNX graph files (no weights) used by WebGPU Chat (source).

Each file is the original model graph with one change: every com.microsoft.MatMulNBits node gets accuracy_level = 4. Nothing else is modified (same nodes, same inputs and outputs, same external-data references). With it, ONNX Runtime's WebGPU backend may use its int8 DP4A matmul path when processing prompts (4 or more tokens at once) on GPUs that expose subgroups. Token-by-token generation is unchanged.

The weights are not here: they are loaded from the original repositories, at the revisions below.

File Source graph Weights
qwen3.5-4b-opt/decoder_model_merged_q4f16.onnx onnx-community/Qwen3.5-4B-ONNX-OPT@bb07226 same repo and revision
gemma-4-12b-qat/decoder_model_merged_q4f16.onnx james58899/gemma-4-12B-it-qat-ONNX@60555e6 (QAT Q4_0 weights, fused GroupQueryAttention) same repo and revision
gemma-4-12b/decoder_model_merged_q4f16.onnx AMINOXDEV/gemma-4-12B-it-ONNX@90e0a27 same repo and revision

Licences follow the original models: Qwen3.5 (Apache 2.0) and Gemma 4 (Apache 2.0).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support