Half of all MoE experts removed from Qwen3-Coder-Next using GOBA-AI-Labs' proprietary calibration-based expert optimization, achieving extreme compression with minimal quality loss.

Inference Engine: This model uses layer-adaptive pruning (different expert counts per layer) and requires moe-stream for inference. llama.cpp does not currently support the experts_per_layer metadata format.

Model Details

Property	Value
Base Model	Qwen/Qwen3-Coder-Next
Architecture	Hybrid MoE (DeltaNet + Attention)
Original Size	45 GB (Q4_K_M)
Pruned Size	24.4 GB (Q4_K_M)
Experts per Layer	Layer-adaptive (226–259, avg ~250, from 512)
MoE Layers	48
Routing	Top-8
Quantization	Q4_K_M
Inference Engine	moe-stream (required)
License	Apache 2.0

Benchmark Results

Benchmark	Original (512 experts)	50% Pruned (~256 experts)	Delta
MMLU (0-shot, 100Q)	77%	72%	-5pp
HumanEval (50Q)	74%	72%	-2pp
LCB Easy (pass@1, 30Q)	—	83.3%	—

93.5% of original MMLU quality retained with 50% of all experts removed.

Size Comparison

Metric	Original	50% Pruned	Savings
File Size (Q4_K_M)	45 GB	24.4 GB	-45.8%
Total Experts	24,576	12,015	-51.1%
Layers	48	48	—

Why This Matters

45 GB → 24 GB: The original model requires 48+ GB RAM. This pruned version fits in 24 GB, making it accessible on consumer hardware
Outperforms Q2 quantization: At similar size (~24 GB), Q2 quantization typically degrades quality by 15-20pp. Our expert pruning loses only 5pp
Expert pruning > aggressive quantization: Removing redundant computation paths preserves model capability better than reducing numerical precision

Usage

This model requires moe-stream for inference due to its layer-adaptive expert structure.

Install

bash

git clone https://github.com/GOBA-AI-Labs/moe-stream
cd moe-stream
cargo build --release --features metal,accelerate

# Download model
huggingface-cli download goba-ai-labs/PrunedHub-Qwen3-Coder-Next-50pct \
  --local-dir models/

CLI Inference

bash

# Text generation
./target/release/moe-stream models/PrunedHub-Qwen3-Coder-Next-50pct-Q4_K_M.gguf 512 \
  --prompt "def fibonacci(n):" --stream \
  --preload-gates --preload-attn

OpenAI-Compatible HTTP Server

bash

# Start server
./target/release/moe-stream-server \
  --model models/PrunedHub-Qwen3-Coder-Next-50pct-Q4_K_M.gguf --port 11434

# Test with curl
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"local","messages":[{"role":"user","content":"Write a Python function to sort a linked list"}],"stream":true}'

Python

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="unused")
response = client.chat.completions.create(
    model="local",
    messages=[{"role": "user", "content": "Implement binary search in Rust"}],
    stream=True
)
for chunk in response:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

Why not llama.cpp?

This model uses layer-adaptive pruning, meaning each layer retains a different number of experts. The per-layer expert counts are stored in the experts_per_layer GGUF metadata field, which llama.cpp does not currently support. moe-stream reads this metadata and correctly routes tokens to the available experts in each layer.

Methodology

Calibration-based importance scoring: Expert importance is measured through actual inference behavior on diverse workloads (academic text, code, mathematics), not just weight magnitude. This is critical at 50% pruning where static analysis would cause severe quality degradation
Layer-adaptive expert allocation: Each of the 48 MoE layers retains a dynamically determined number of experts. Some layers are more sensitive to pruning than others — adaptive allocation preserves quality where it matters most
Expert pruning vs quantization: At ~24 GB, aggressive quantization (Q2/Q3) would degrade all computations uniformly. Expert pruning instead removes entire redundant computation paths while keeping the remaining experts at full Q4 precision, preserving reasoning capability
Cross-architecture validated: The same methodology has been validated on GPT-OSS-20B (lossless at 12.5% pruning) and Qwen3-30B-A3B (near-lossless at 20% pruning), demonstrating generalization across MoE architectures

Inference Engine: moe-stream

moe-stream is a Rust-based MoE inference engine by GOBA-AI-Labs.

Feature	Details
Inference Modes	GPU Resident / GPU Hybrid / SSD Streaming (auto-selected)
GPU Support	Apple Metal / NVIDIA CUDA
Quantization	Q2K-Q8K, MXFP4, F16, F32 (13 formats)
API	OpenAI-compatible HTTP / JSONL / MCP
Special	Q4 Quantized MatMul (+79% speedup), Dynamic K

Citation

bibtex

@misc{goba-ai-labs-prunedhub-qwen3-coder-next-50pct,
  title={PrunedHub Qwen3-Coder-Next-50pct: Extreme MoE Compression via Expert Pruning},
  author={GOBA-AI-Labs},
  year={2026},
  url={https://huggingface.co/GOBA-AI-Labs/PrunedHub-Qwen3-Coder-Next-50pct}
}

GOBA-AI-Labs/PrunedHub-Qwen3-Coder-Next-50pct

PrunedHub Qwen3-Coder-Next-50pct — MoE-Stream Edition

Model Details

Benchmark Results

Size Comparison

Why This Matters

Usage

Install

CLI Inference

OpenAI-Compatible HTTP Server

Python

Why not llama.cpp?

Methodology

Inference Engine: moe-stream

Citation

Links

GOBA-AI-Labs/
PrunedHub-Qwen3-Coder-Next-50pct