{
  "_comment": "VIDEO 035 part 1 (2026-08-28 night): GLM-5.3-Flash UD-IQ1_S, reasoning_effort LOW (GLM has low/high/max only; low keeps a 10 tok/s model inside an overnight). Unsloth branch glm5next/upstream build (Vulkan), flash attention OFF per PR #27754. Samplers = model card (1.0 / 0.95). Context 65536 to leave memory for the 2-bit leg.",
  "port": 8899,
  "output_dir": "runs/glm-035",
  "max_tokens": 120000,
  "seed": 42,
  "request_timeout_minutes": 240,
  "tests": [
    "01-blockfall",
    "03-eruption",
    "04-the-ledger",
    "05-blind-artist"
  ],
  "contestants": [
    {
      "id": "glm53flash-iq1s-low",
      "label": "GLM-5.3-Flash UD-IQ1_S (93.1 GB) reasoning LOW",
      "type": "llamacpp",
      "repo": "unsloth/GLM-5.3-Flash-GGUF",
      "variant": "UD-IQ1_S",
      "vision": false,
      "launch_chat_template_kwargs": {
        "reasoning_effort": "low",
        "clear_thinking": true
      },
      "sampling": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 0,
        "min_p": 0,
        "presence_penalty": 0,
        "repeat_penalty": 1.0
      },
      "max_seq_length": 65536,
      "binary": "~/private/jose-yt/tools/llama.cpp-glm5next/build/bin/llama-server",
      "extra_launch_args": [
        "-fa",
        "off"
      ]
    }
  ]
}