At a glance

Which variant should I pick?

20% expert-pruned version of Qwen/Qwen3-Coder-Next using Cerebras REAP (Router-weighted Expert Activation Pruning).

REAP removes 20% of MoE experts (102 of 512) while preserving the model's routing behavior and output quality. The active parameter count per token is unchanged since the router still selects 10 experts per token from the remaining pool. This yields a ~14% reduction in total disk/memory footprint with minimal quality loss.

Method

REAP (ICLR 2026) prunes Mixture-of-Experts models by scoring expert importance using:

1. Router gate values -- how often and how strongly the router selects each expert
2. Expert activation norms -- magnitude of each expert's output contribution
3. Frequency-weighted saliency -- combining routing frequency with activation importance
4. Router logit renormalization -- maintains output distribution after expert removal
5. Layerwise application -- independent per-layer pruning decisions for stability

Calibration Dataset

22,000 samples (no-refusal subset: 21,000), packed to 16,384 token sequences:

Total tokens observed: ~90.5M across 6,391 packed sequences.

Pruning Configuration

Benchmark Results

10-task lm-eval suite, 200 samples per task, tensor_parallel_size=4, vLLM eager mode:

Aggregate:

• Overall average: 66.7% -> 64.6% (-2.1 pts)
• Reasoning average: 71.4% -> 70.5% (-0.9 pts)
• Math average: 47.8% -> 41.0% (-6.8 pts)

Architecture

Qwen3-Coder-Next uses a hybrid linear/full attention architecture with 48 layers:

• Full attention every 4th layer (12 layers)
• Linear attention for remaining layers (36 layers)
• MoE FFN with 410 remaining experts per layer, 10 active per token
• Shared expert (intermediate size 512) in every layer
• Context window: 262,144 tokens
• Vocab size: 151,936

Usage

Transformers

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xSero/Qwen3-Coder-64B"

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=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

vLLM

bash
Copy code
vllm serve 0xSero/Qwen3-Coder-64B \
    --tensor-parallel-size 4 \
    --enforce-eager \
    --gpu-memory-utilization 0.9 \
    --max-model-len 32768

Reproducing

bash
Copy code
git clone https://github.com/cerebras/reap
cd reap

python -m reap.layerwise_prune \
    --model-name Qwen/Qwen3-Coder-Next \
    --dataset-name combined \
    --compression-ratio 0.20 \
    --prune-method reap \
    --seed 42 \
    --renormalize_router_weights true \
    --batch_size 8 \
    --batches_per_category 128

Links

• REAP paper: arxiv.org/abs/2510.13999
• REAP code: github.com/cerebras/reap
• Cerebras REAP collection: huggingface.co/collections/cerebras/cerebras-reap
• Base model: Qwen/Qwen3-Coder-Next
• 30% pruned variant: 0xSero/Qwen3-Coder-57B

License & citation

License inherited from the base model.

bibtex
Copy code
@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.