You can deploy this model and start earning money today!

Views

No views yet

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

Qwen3.5-28B

REAP-pruned Qwen/Qwen3.5-35B-A3B.

At a glance


Base model	Qwen/Qwen3.5-35B-A3B
Format	BF16
Total params	28B
Active / token	3B
Experts / layer	—
Layers	—
Hidden size	—
Context	—
On-disk size	57 GB

Which variant should I pick?

Variant	Format	Link
`Qwen3.5-264B`	BF16	link
`Qwen3.5-264B-FP8`	FP8	link
`Qwen3.5-264B-W4A16`	W4A16	link
`Qwen3.5-28B` (this)	BF16	link
`Qwen3.5-35B-EXL3-4bpw`	EXL3-4bpw	link
`Qwen3.5-76B`	BF16	link
`Qwen3.5-76B-GGUF`	GGUF	link
`Qwen3.5-88B`	BF16	link
`Qwen3.5-99B`	BF16	link
`Qwen3.5-99B-GGUF`	GGUF	link

This is a 20% expert-pruned version of Qwen/Qwen3.5-35B-A3B, produced using the REAP (Router-weighted Expert Activation Pruning) method from the paper "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression" (ICLR 2026).

The model retains 205 of 256 experts (~80% of original experts) while remaining competitive on standard benchmarks.

What We Did

Method: REAP Layerwise Pruning

REAP prunes MoE experts by scoring each expert's importance using a combination of:

Router gate-values — how often and how strongly the router selects each expert
Expert activation norms — the magnitude of each expert's output contribution

Router logit weights are renormalized to sum to 1 after pruning (critical for maintaining output scale). Pruning is applied layer-by-layer (layerwise mode).

Calibration Data

Observations were collected over a mixed calibration dataset of 1,000 samples per category:

theblackcat102/evol-codealpaca-v1 (250 samples)
open-r1/Mixture-of-Thoughts[code] (250 samples)
open-r1/Mixture-of-Thoughts[math] (250 samples)
open-r1/Mixture-of-Thoughts[science] (250 samples)

Max sequence length: 4096 tokens. Angular distance measure for expert similarity.

Pruning Config

Parameter	Value
Compression ratio	20% (51 experts removed)
Original experts	256
Remaining experts	205
Pruning method	REAP
Router weight renormalization	✓
Seed	42
Calibration samples	1,000 total

Benchmark Results

All evaluations run with vLLM (tensor-parallel across 8x RTX 3090), greedy decoding, 0-shot.

Coding (EvalPlus, greedy, 0-shot)

Benchmark	Original	Pruned (20%)	Delta
HumanEval (pass@1)	76.2%	73.2%	-3.0%
HumanEval+ (pass@1)	72.0%	70.1%	-1.9%

Multiple Choice / Reasoning (lm-eval, 0-shot, 250 samples/task)

Benchmark	Original	Pruned (20%)	Delta
MMLU	84.34%	80.89%	-3.45%
MMLU - Humanities	82.40%	76.35%	-6.05%
MMLU - Social Sciences	90.04%	88.38%	-1.66%
MMLU - STEM	81.46%	78.88%	-2.58%
MMLU - Other	84.52%	81.05%	-3.47%
ARC-Challenge	60.00%	60.40%	+0.40%
ARC-Easy	84.00%	83.20%	-0.80%
BoolQ	88.00%	89.20%	+1.20%
HellaSwag (norm)	76.40%	75.60%	-0.80%
OpenBookQA (norm)	45.20%	47.20%	+2.00%
RTE	81.20%	82.00%	+0.80%
WinoGrande	77.20%	76.80%	-0.40%

Perplexity (WikiText-2, 10k tokens, llama.cpp)

Model	PPL
Original (256 experts)	6.83
Pruned 20% (205 experts)	9.51

Throughput (4x RTX 3090, TP=4, vLLM, enforce_eager)

Batch Size	Original tok/s	Pruned tok/s	Speedup
1	12.3	12.5	1.02x
4	37.0	36.0	0.97x
8	74.4	70.3	0.95x
16	89.3	86.0	0.96x

Note: throughput speedup is minimal at this compression level with current vLLM routing overhead. The primary benefit is reduced VRAM footprint.

Memory Footprint

Model	Size	Shards
Original	~71 GB (bf16)	14 safetensors
Pruned 20%	~53 GB (bf16)	2 safetensors

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xSero/Qwen3.5-35B-A3B-REAP-20pct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

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))

With vLLM:

bash

vllm serve 0xSero/Qwen3.5-35B-A3B-REAP-20pct \
    --tensor-parallel-size 4 \
    --gpu-memory-utilization 0.9 \
    --max-model-len 32768

Reproducing

bash

git clone https://github.com/cerebras/reap
cd reap
bash scripts/build.sh

python -m reap.layerwise_prune \
    --model_name Qwen/Qwen3.5-35B-A3B \
    --dataset_name "theblackcat102/evol-codealpaca-v1:250,open-r1/Mixture-of-Thoughts[code]:250,open-r1/Mixture-of-Thoughts[math]:250,open-r1/Mixture-of-Thoughts[science]:250" \
    --compression_ratio 0.20 \
    --prune_method reap \
    --seed 42 \
    --renormalize_router_weights true

License & citation

License inherited from the base model.

bibtex

@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}
}

0xSero/
Qwen3.5-28B

Qwen3.5-28B

At a glance

Which variant should I pick?

What We Did

Method: REAP Layerwise Pruning

Calibration Data

Pruning Config

Benchmark Results

Coding (EvalPlus, greedy, 0-shot)

Multiple Choice / Reasoning (lm-eval, 0-shot, 250 samples/task)

Perplexity (WikiText-2, 10k tokens, llama.cpp)

Throughput (4x RTX 3090, TP=4, vLLM, enforce_eager)

Memory Footprint

Usage

Reproducing

License & citation

Sponsors

0xSero/Qwen3.5-28B

Qwen3.5-28B

At a glance

Which variant should I pick?

What We Did

Method: REAP Layerwise Pruning

Calibration Data

Pruning Config

Benchmark Results

Coding (EvalPlus, greedy, 0-shot)

Multiple Choice / Reasoning (lm-eval, 0-shot, 250 samples/task)

Perplexity (WikiText-2, 10k tokens, llama.cpp)

Throughput (4x RTX 3090, TP=4, vLLM, enforce_eager)

Memory Footprint

Usage

Reproducing

License & citation

Sponsors

0xSero/
Qwen3.5-28B