Wanda channel pruning: FFN intermediate 10752 → 7168 (uniform across all 30 layers), joint swiglu importance (w1/w3 rows + w2 cols, one shared mask) from multimodal calibration (wikitext-2 + COCO), equal text/image weight.
Distillation recovery: r32 LoRA on all FFN linears trained to match the base model's distribution (KL over the base's top-64 logits + small NLL anchor) on 512 rows disjoint from evaluation data (two low-LR passes).
Evaluation: our statistical integrity harness vs the base.

Quality vs base (quick suite, our harness)

Axis	Value
text PPL ratio (e^ΔNLL)	0.85 (≤1.05 gate: PASS)
mm PPL ratio	0.81 (≤1.10 gate: PASS)
text kld_t64	0.257 (base = 0)
mm kld_t64	0.260
generation chrF (text/mm)	0.48 / 0.50
integrity index (record-only)	41.7 / 100

Deeper cut than the 8192 sibling: slightly more distribution drift, still no word-prediction drift (PPL gates pass).

Load

python

from transformers import AutoModelForImageTextToText
model = AutoModelForImageTextToText.from_pretrained(
    "konic-labs/LFM2.5-VL-3B-ffn7168", dtype="bfloat16")

License & provenance

Derivative of LiquidAI/LFM2.5-VL-3B (LFM Open License v1.0 — license text retained; modifications per this card).
Pipeline and full reports: https://github.com/konic-labs/lfm25-vl-3b-lab
Siblings: LFM2.5-VL-3B-ffn8192, LFM2.5-VL-3B-ffn8192-int4.