{
  "experiment": "GLM-5.3 Flash vs GLM-5.3 Full 743B interactive solar system",
  "date": "2026-08-29",
  "thinking_token_note": "Tokenizer-counted from each preserved reasoning stream with the exact GLM tokenizer. Native usage did not split reasoning tokens.",
  "flash": {
    "deployment": "GLM-5.3 Flash NVFP4, TP2 on 2 DGX Sparks, DFlash2 K7",
    "thinking_tokens_tokenizer_counted": 29945,
    "prompt_tokens": 422,
    "completion_tokens": 46418,
    "total_tokens": 46840,
    "wall_seconds": 1869.3083126661368,
    "completion_tokens_per_second_after_first_delta": 24.93679918007907,
    "finish_reason": "stop",
    "visual_result": "Native visual fail_major; public live view is a disclosed two-line display repair."
  },
  "full": {
    "deployment": "GLM-5.3 Full 743B Int4/Int8Mix, TP4 on 4 DGX Sparks, in-checkpoint MTP K5",
    "thinking_tokens_tokenizer_counted": 82364,
    "prompt_tokens": 422,
    "completion_tokens": 105787,
    "total_tokens": 106209,
    "wall_seconds": 4916.743224167032,
    "completion_tokens_per_second_after_first_delta": 21.522897190970177,
    "finish_reason": "stop",
    "visual_result": "Unmodified output passed structural, runtime, offline, and motion validation."
  }
}