Card: Usage now gives the runner settings that load these decode-only bundles (llm-runner sequential engine with one-token prefill, CoreAIKit ChatSession), checked 2026-10-02; drop the stock-pipelined-engine claim
gemma4_12b_qat_decode_int4linsym_msdpa_g8: re-save with coreai-core 1.0.0b2 (strip debug locations; weights unchanged). The 0.4.0-era IR refused AIModel.load on every OS 27 build since beta 2, still on the RC 26A428 (2026-09-14); the re-saved bundle loads. Recovery: coreai-model-zoo conversion/recovery/{strip_b1,resave_b2}.py
gemma4_12b_qat_decode_int8lin_msdpa_g8: re-save with coreai-core 1.0.0b2 (strip debug locations; weights unchanged). The 0.4.0-era IR refused AIModel.load on every OS 27 build since beta 2, still on the RC 26A428 (2026-09-14); the re-saved bundle loads. Recovery: coreai-model-zoo conversion/recovery/{strip_b1,resave_b2}.py
The bundle shipped the template as it stood at export time. Google's revision fixes tool-calling loops, turn closures and thinking content-ordering, and this bundle's graph is unaffected — only prompt formatting changes.
Found by conversion/zoo_verify.py in github.com/john-rocky/coreai-model-zoo.
