Qwen3.8-Flash-Next-AZ-Focus on one DGX Spark: receipts

Serving (one Spark per model, identical): Mia's single-Spark TensorFold recipe
  https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold
  base:     MODEL_ID=azampatti/Qwen3.8-Flash-Next-125B-A5B-INT4-AutoRound (recipe default)
  AZ-Focus: MODEL_ID=AxonZeta/Qwen3.8-Flash-Next-AZ-Focus-INT4 MODEL_REVISION=5d2f89941ae201fa3458515a6ab06dd51c25e60b
  both with MAX_TOKENS=131072 in the recipe's .env (128k output limit; no answer was cut off).

Time to complete (ttc-*.json): ttc.py, the 60 questions in ttc-questions.json (30 MMLU-Pro, 30 BIG-Bench Hard),
  python3 ttc.py --base-url http://HOST:8888 --model Qwen3.8-Flash-Next --label L --out OUT.json \
    --concurrency 1|4 --effort xhigh --max-tokens 131072
  (xhigh is Qwen3.8's default effort; sampling temperature 1.0, top_p 0.95, seed per question.)

Benchmarks (bench-base.jsonl, bench-az-focus.jsonl): one line per answer, 1,298 per model: MMLU-Pro (500), BIG-Bench
  Hard (540), GPQA Diamond (198), AIME 2025 (30 questions x 2 seeds); default effort, 128k output limit, same sampling.
  Fields: id, suite, sample, correct, reasoning_tokens, completion_tokens, finish.

RigMark (rigmark-*.json receipts and *.card.txt cards): github.com/alexellis/rigmark at 770c3106 (protocol 1.3),
  one Spark, the same recipe settings as above (MAX_TOKENS=131072, VISION=0), base then AZ-Focus on the same Spark:
  ./rigmark run --base-url http://HOST:8888 --model Qwen3.8-Flash-Next --label L --comparison-id ID --metadata META.json \
    --extra-body '{"chat_template_kwargs": {"reasoning_effort": "xhigh"}}' --decode-tokens 32768 --runs 3
  and at low effort with RigMark's defaults (--extra-body '{"chat_template_kwargs": {"reasoning_effort": "low"}}').
  Each JSON records the checkpoint revision, image and settings; temperature 0.