# Qwen3.8-Flash-Next bf16 reference logits

Ground truth for Splash quality checks, from
`dev/benchmarks/qwen4exp/reference_logits.py` (full-precision checkpoint,
float32 on CPU). Each `<name>.reference.bin` is bf16 logits [tokens x 248320]
for `<name>.passage.json`.

- code: 1,495-token code prompt; short: 65 tokens; long4k: first 4,096 tokens
  of the long prompt (crosses the 2,048-token sparse-indexer budget).
- code.plan.bin: the reference run with the current package's quantization
  simulated (`--quant experts:4,router:8,shared:4,attn:8,gdn_in:8,gdn_ab:8,
  gdn_out:8,ple:4,head:8,embed:8`): 90.2% same pick, KL 0.122.

Score an engine dump:
    SPLASH_DUMP_PREFILL_LOGITS=/tmp/x.bin build/engine-tests/generate-sample \
      build/splash.metallib ~/models/qwen38-flash-next-splash 1 "<ids>"
    python dev/benchmarks/qwen4exp/compare_logits.py --passages code.passage.json \
      --name code --reference code.reference.bin --engine /tmp/x.bin [--begin N --end M]
(Use ~/vllm-mlx-env-3.12/bin/python for the reference and compare scripts.)