# Released Checkpoints

The checkpoints below are the model artifacts used for the reported best EchoLoc Thinker+Talker evaluation. They are stored in this repository under `checkpoints/` and are separate from the source-only `MANIFEST.tsv`.

## Thinker

- Repository path: `checkpoints/thinker/`
- Training artifact: Iter006 `latest_final_model`
- Format: PEFT LoRA adapter with tokenizer and processor assets
- Base model: `Qwen/Qwen3-Omni-30B-A3B-Instruct`
- Main file: `adapter_model.safetensors`
- Size: `4,642,233,208` bytes
- SHA-256: `cf716ae5bb40a49ef75f642229a0e7ca74e381925e21c66f8d161aa128514ac2`

## Talker

- Repository path: `checkpoints/talker/`
- Training artifact: `qwen3omni_talker_vstyle_46000`
- Format: Talker state dictionary with tokenizer assets
- Main file: `pytorch_model.bin`
- Size: `3,833,490,035` bytes
- SHA-256: `cd1e07334a94b59f827f55c6805b80a3ac8e0ed9b9c92217590ede712783b335`

The Talker loader first initializes the merged Qwen3-TTS Base/VoiceDesign origin and then loads this state dictionary into the inner Talker. See `checkpoints/talker/README.md` and `RUNBOOK.md` for the exact commands.