[!Warning] The conversion scripts were written by Claude, and the conversion was run with them. Check it yourself before use; use it at your own risk.
This is a text-only ONNX conversion of
google/gemma-4-12B-it-qat-q4_0-unquantized
(revision b6ed86275a6a5735884e208bfed95b445a684ca2). The weights are 4-bit (block 32), taken from the QAT grid of
that checkpoint.
onnx/embed_tokens_{q4,q4f16}.onnx: the token embedding.onnx/decoder_model_merged_{q4,q4f16}.onnx: the decoder, with a KV cache.q4: fp32 activations and KV cache.q4f16: fp16 activations and KV cache.config.json: the text part of the source config (Gemma4UnifiedForCausalLM). There is no vision or audio encoder.scripts/: the conversion scripts that produced the model (python scripts/export.py <checkpoint dir> <out dir> [--fp32]; --fp32 gives q4).Apache 2.0, same as Gemma 4.
onnx/decoder_model_merged_q4.onnx 539.5 kBonnx/decoder_model_merged_q4f16.onnx 455.3 kBonnx/embed_tokens_q4f16.onnx 0.6 kBonnx/embed_tokens_q4.onnx 0.6 kB有文件?提交种子。
提交种子