{
  "_comment": "VIDEO 034 (staged 2026-08-26 night): Qwen 3.8-Flash-Next first benchmarks. Protocol v1.1 four tests, THINKING ON at reasoning_effort medium with the card's thinking samplers (1.0/0.95/20/0), so every leg is directly comparable to runs/thinking-levels-q4/q4-think-medium (Qwen 3.8-27B UD-Q4_K_XL, same tests, same samplers, same effort). Run order = priority: UD-IQ1_M (75 GB, the one Unsloth says fits) first, then UD-Q2_K_XL (79 GB, 'does it even load on 128 GB' is itself a finding), then the optional 27B BF16 medium leg last (slow, may run into the day; if it DNFs the Q4-medium data is the comparison). Limits match 030: 120k output budget inside 131072 ctx, 240-min timeout as a hang guard only. Speculative/MTP: launcher default, logged from the server banner, same as the 27B runs. A contestant that fails to load is logged as CONTESTANT FAILED and the night continues. BACKEND NOTE: the Flash-Next legs run on a from-source llama.cpp build of ggml-org PR #27742 (Vulkan/RADV on the 8060S) served directly by llama-server; the 27B legs ran on Unsloth's ROCm prebuilt. Quality is comparable (same prompts, samplers, seed, effort); tok/s across the two backends is reported with the backend named, never as a straight ratio.",
  "port": 8899,
  "output_dir": "runs/flash-next-034",
  "max_tokens": 120000,
  "seed": 42,
  "request_timeout_minutes": 240,
  "tests": [
    "01-blockfall",
    "03-eruption",
    "04-the-ledger",
    "05-blind-artist"
  ],
  "contestants": [
    {
      "id": "flashnext-iq1m-think-medium",
      "label": "Qwen 3.8-Flash-Next UD-IQ1_M (74.5 GB) THINKING medium (card samplers)",
      "type": "llamacpp",
      "repo": "unsloth/Qwen3.8-Flash-Next-GGUF",
      "variant": "UD-IQ1_M",
      "vision": false,
      "launch_chat_template_kwargs": {
        "enable_thinking": true,
        "reasoning_effort": "medium"
      },
      "sampling": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0,
        "presence_penalty": 0,
        "repeat_penalty": 1.0
      },
      "max_seq_length": 131072,
      "binary": "~/private/jose-yt/tools/llama.cpp-qwen4exp/build/bin/llama-server",
      "extra_launch_args": [
        "--flash-attn",
        "on"
      ]
    },
    {
      "id": "flashnext-q2kxl-think-medium",
      "label": "Qwen 3.8-Flash-Next UD-Q2_K_XL (78.9 GB) THINKING medium (card samplers)",
      "type": "llamacpp",
      "repo": "unsloth/Qwen3.8-Flash-Next-GGUF",
      "variant": "UD-Q2_K_XL",
      "vision": false,
      "launch_chat_template_kwargs": {
        "enable_thinking": true,
        "reasoning_effort": "medium"
      },
      "sampling": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0,
        "presence_penalty": 0,
        "repeat_penalty": 1.0
      },
      "max_seq_length": 131072,
      "binary": "~/private/jose-yt/tools/llama.cpp-qwen4exp/build/bin/llama-server",
      "extra_launch_args": [
        "--flash-attn",
        "on"
      ]
    },
    {
      "id": "qwen38-27b-bf16-think-medium",
      "label": "Qwen 3.8-27B BF16 (54.7 GB) THINKING medium (card samplers), optional slow leg",
      "type": "unsloth",
      "repo": "unsloth/Qwen3.8-27B-GGUF",
      "variant": "BF16",
      "vision": false,
      "launch_chat_template_kwargs": {
        "enable_thinking": true,
        "reasoning_effort": "medium"
      },
      "sampling": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0,
        "presence_penalty": 0,
        "repeat_penalty": 1.0
      },
      "max_seq_length": 131072
    }
  ]
}