{
  "model_name": "hojo-tts-light-40m",
  "task": "text-to-speech (24kHz, 15 voices), 3 模型流水线 (ax-llm + fine_local + decoder)",
  "route": "ax-llm lm_s8 + fine_local INT8 + decoder_sq SmoothQuant/U16",
  "lm": {
    "type": "Qwen3 10L decode-only (s8 weights, bf16 hidden)",
    "tool": "Pulsar2 7.0 llm_build2, max_context 2176, NPU3",
    "files": "models/lm_s8/qwen3_p8_l0..9_together.axmodel + qwen3_post.axmodel + embed_tokens.bin + config.json + post_config.json + tokenizer.*",
    "sampling": "deterministic greedy (enable_temperature=false)"
  },
  "fine_local": {
    "file": "models/fine_local.axmodel",
    "precision": "INT8",
    "cos": 0.9997
  },
  "decoder": {
    "file": "models/decoder_sq.axmodel",
    "precision": "SmoothQuant+U16",
    "mag": "0.998+",
    "phase": "0.95-0.97"
  },
  "sdk": {
    "bin": [
      "bin/hojo_tts_cpp",
      "bin/tts_driver"
    ],
    "usage": "hojo_tts_cpp <lm_dir> <embeds.bin> <num_tokens> <speech_end_id> <max_new_tokens> <fine_local.axmodel> <decoder.axmodel> <embed_tokens.bin> <speaker_vecs.bin> <voice_idx> <id2code.bin> <out.wav>"
  }
}