kode-local-models

ONNX exports used by Kode's local router. Kode downloads these with kode setup, verifies every file by sha256, and runs them in-process on ONNX Runtime.

Folder	Source	What it is
laya-multilingual/	convaiinnovations/laya (multilingual/ checkpoint)	Laya decision model (mmBERT-base encoder + decision head), fp32. Inputs input_ids, attention_mask, marker_pos, marker_mask, qtype (int64); output logits [b, k].
qwen3-reranker-0.6b/	Qwen/Qwen3-Reranker-0.6B	Slim reranker graph, fp32: output yes_no [b, 2] = logits for "no"/"yes" at the last position. Weights in model.onnx.data.

Both upstream models are Apache-2.0; these are unmodified weights in a different serialization, redistributed under the same license.

Built with scripts/local-models/build_laya.py and build_reranker.py in the Kode repo. Each export was checked against the upstream reference code before upload (max logit difference: Laya 6.9e-6, reranker 5.2e-7).