Gemma-4-19B

[!TIP] Support this work → · X · GitHub · REAP paper · Cerebras REAP

Gemma-4-19B

REAP-pruned google/gemma-4-26b-a4b-it.

At a glance

Which variant should I pick?

30% expert-pruned version of google/gemma-4-26b-a4b-it using Cerebras REAP (Router-weighted Expert Activation Pruning).

REAP removes 30% of MoE experts (38 of 128 per layer) while preserving the model's routing behavior. The active parameter count per token is unchanged since the router still selects 8 experts per token from the remaining pool. This yields a ~31% reduction in total disk/memory footprint.

How This Model Was Made

Step 1: Calibration (Activation Observation)

We ran the full Gemma 4 26B-A4B-it model over a curated calibration dataset to record expert activation patterns across all layers.

Calibration dataset: 22,000 samples drawn from 12 sources:

Step 2: REAP Pruning

The lowest-scoring 30% of experts (38 per layer) are removed based on combined router gate values, activation norms, and frequency-weighted saliency. Router logits are renormalized post-pruning.

Pruning Configuration

Benchmark Results

Accuracy (generative, 0-shot, 50 samples/task, thinking enabled, vLLM 0.19, 4x RTX 3090)

Evaluated using lm-eval generative tasks with --apply_chat_template and think_end_token=<channel|> to properly handle Gemma 4's thinking mode.

* Tasks with extraction failures. Real accuracy likely higher.

Summary: At 30% pruning, the model retains strong performance on elementary math (88%) and coding (68% college CS). Knowledge-intensive tasks like philosophy and world religions see moderate drops. The 30% variant trades ~18 pts on philosophy and ~42 pts on world religions for a 31% smaller model.

Robustness: Generation Quality (14 prompts, temp=0.7, top_p=0.9, max 2048 tokens)

Zero looping, zero collapse across coding, math, and philosophy. The REAP 0.30 model actually outperforms the original on long-context (0% vs 50% loop rate). Generation quality is fully preserved at 30% compression.

Architecture

Gemma 4 uses a hybrid sliding/full attention MoE architecture:

• 30 transformer layers
• Sliding attention (window=1024) for 25 layers, full attention every 6th layer
• MoE FFN with 90 remaining experts per layer (originally 128), 8 active per token
• Thinking model -- uses <|channel>thought / <|channel>response channels
• Multimodal -- supports text and vision inputs
• Context window: 262,144 tokens
• Vocab size: 262,144

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xSero/Gemma-4-19B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True)

messages = [{"role": "user", "content": "Write a quicksort in Python."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

vLLM

pip install vllm>=0.19 transformers>=5.0

vllm serve 0xSero/Gemma-4-19B \
    --tensor-parallel-size 2 \
    --enforce-eager \
    --gpu-memory-utilization 0.9 \
    --max-model-len 16384 \
    --trust-remote-code

Links

• REAP paper: arxiv.org/abs/2510.13999
• REAP code: github.com/cerebras/reap
• 20% pruned variant: 0xSero/Gemma-4-21B
• Base model: google/gemma-4-26b-a4b-it

License & citation

License inherited from the base model.

@misc{lasby2025reap,
  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
  year   = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}

Sponsors

Made possible by NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle.