flint-imatrix / README.md
hermitdave's picture
coverage p05 13662 (corrected capture)
052c464 verified
|
Raw History Blame Contribute Delete
1.25 kB
---
license: other
base_model:
- RecursiveSI/flint
tags:
- rco
- mixed-precision
- quantized
- imatrix
library_name: mlx
---
# Flint — RCO imatrix
Importance-matrix calibration data for [RecursiveSI/flint](https://huggingface.co/RecursiveSI/flint)
(39.8 B total / ~2.5 B active, 49 layers: 2 dense + 47 MoE, 256 routed experts
top-6 plus a shared expert).
| | |
|---|---|
| tensors | 36 487 (36 096 per-expert) |
| chunks | 512 x 2048 tokens |
| calibration | FineWeb-Edu, checkpoint's own tokenizer (roundtrip 1.000) |
| coverage | 36 093 / 36 096 expert-projection keys active; 1 expert never routed (3 keys) carries the layer-pooled fallback; p05 = 13 662 tokens/expert |
| expert keys | per-expert (up / down / gate) |
Unlike Qwen3.6 — whose checkpoint stores routed experts **fused**
(`mlp.experts.gate_up_proj`, no expert index) and therefore cannot be
addressed per expert — Flint stores `mlp.experts.{i}.{up,down,gate}_proj`
separately, so this imatrix carries a distinct statistic for each of the
256 experts in every MoE layer.
Captured with `scripts/capture_imatrix_streaming.py --model-type flint`
from the RCO MLX pipeline. Quantization workflow assisted by
[Hermes Agent](https://hermes-agent.nousresearch.com).