mrysh/gemma-4-e4b-it-prune_depth_10

gemma-4-e4b-it - prune_depth_10

Pruned version of google/gemma-4-E4B-it made for a pruning benchmark (4 models x 8 pruning variants). The benchmark compares the variants with each other and with quantization on BFCL, hiring-bias and MedQA, run with vLLM.

• method: depth, level 0.1
• library: nvidia-modelopt 0.46.0rc1
• size on disk: 14.649 GB
• reduction of the language model: 9.2% of its parameters (of the whole checkpoint: 8.6%)

Depth pruning with NVIDIA ModelOpt Minitron: 42 -> 38 layers (10.0% of the layers requested), the width is unchanged. Importance of the layers was measured on a calibration set of 512 mixed samples.

The vision / audio blocks of the base model, if it has them, are kept unchanged and only the language part is pruned. The multi-token-prediction head of Qwen3.5 is dropped by the pruning tool.

No distillation or fine-tuning was done after pruning, so expect a quality drop against the base model; the benchmark measures that drop. In the benchmark it is evaluated as a text model (vLLM, thinking off). Same license as the base model; see the base model card.