AtomicChat/Ling-3.0-flash-GGUF-metrics

Ling-3.0-flash — quantization metrics

Everything measured while building the GGUF line for `inclusionAI/Ling-3.0-flash`: raw logs, per-rung numbers and the importance matrix statistics. Published so the quant table can be checked rather than trusted.

Quants live in `AtomicChat/Ling-3.0-flash-GGUF`.

Layout

metrics/
  grid-table.json          per rung: size, bpw, mean/99% KLD, top-1 agreement
  kld-results.json         raw parser output of every KL divergence run
  quant-results.json       quantize timings and produced sizes
  imatrix-coverage.json    how many of the 512 experts the calibration corpus reached
logs/
  imatrix/                 the full bf16 imatrix run and its statistics
  kld/                     one log per measured rung, plus the baseline
  quant/                   one log per quantization, with the applied tensor overrides
  build/                   how the binaries were built
runbook.md                 the commands, start to finish

Method

• —Baseline: BF16 GGUF, verified bit-exact against the safetensors release (877 of 917 tensors byte-identical, the other 40 are the MoE routers stored as F32, a lossless widening).
• —Importance matrix: collected on the BF16 model, not on a quantized proxy. 522 chunks of 4096 tokens (2.14M tokens) from `AtomicChat/calib-corpora`. Expert coverage 99.98%: 12 of 61440 expert slots never activated.
• —Evaluation: held-out text that never entered the calibration corpus (eval_neutral + calib_heldout), 92 chunks at n_ctx 4096.
• —Control: the baseline measured against itself gives KLD 0.00000 ± 0.00000, top-1 99.997%, so the numbers below are the models' own error and not the harness.
• —Hardware: every KL divergence run on the same 4x RTX PRO 6000 Blackwell box. Cross-GPU comparisons are not meaningful, so hardware is stated with each speed figure.

Reading the numbers

This model floors at a higher KL divergence than a dense model of similar size: even Q8_0 sits at 0.0196. With 8 of 512 experts active per token, a small weight perturbation sometimes changes which expert runs, and the output moves discontinuously. What matters is the increase above that floor, not the absolute value.