Preserve the original Final checkpoint and native complex-valued architecture. Add 264 learned parameter tensors in SafeTensors and 32 saved fixed mixer buffers separately, exposing 498,214 learned parameter elements without counting fixed state. Correct the config to D324/L16/E66/T128/V499 and the logged batch size of 4. Add a strict native loader, original tokenizer implementation, reproducible export and conversion report. Update the card and download/loading examples. All 296 tensors matched byte for byte after strict loading on CPU; no CUDA initialization or inference was performed.
Co-Authored-By: Codex <noreply@openai.com>