CoolFace
Modelpublic

GOBA-AI-Labs/PrunedHub-GPT-OSS-20B-28x

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face
1likes352downloads
Model Card

PrunedHub GPT-OSS-20B-28x

Lossless MoE Expert Pruning — 12.5% of experts removed with zero quality loss.

Pruned from OpenAI GPT-OSS-20B using GOBA-AI-Labs' proprietary expert optimization methodology.

Model Details

PropertyValue
Base Modelopenai/gpt-oss-20b
Total Parameters~17.5B
Active Parameters3.6B per token
Experts per Layer28 (from 32, uniform)
MoE Layers24
RoutingTop-4, sigmoid activation
Context Length128K tokens
QuantizationQ4KM
LicenseApache 2.0

Benchmark Results

BenchmarkOriginal (32 experts)Pruned (28 experts)Delta
MMLU (0-shot, 100Q)78%78%0 pp
GSM8K (0-shot, 50Q)70%*92%—
HumanEval (50Q)80%78%-2 pp
Inference Speed (M4 Pro)~55 tok/s~55 tok/s—

*GSM8K original baseline measured with reasoning budget constraints.

Size Comparison

MetricOriginalPrunedSavings
File Size11.67 GB10.40 GB-10.9%
Experts/Layer3228-12.5%
Total Experts768672-96 experts

Key Features

  • —Lossless compression — Zero quality degradation across all benchmarks
  • —Drop-in replacement — Compatible with llama.cpp (build 7970+)
  • —Fits 16GB RAM — With room for KV cache at 4K context
  • —~55 tok/s on Apple M4 Pro (Metal GPU)

Usage

llama.cpp (recommended)

This model uses uniform expert counts and is fully compatible with llama.cpp:

bash
llama-server -m PrunedHub-GPT-OSS-20B-28x-Q4_K_M.gguf --port 8090 -ngl 99 -c 4096

moe-stream

Also supported by moe-stream, which offers GPU-resident inference and OpenAI-compatible HTTP API:

bash
# CLI inference
moe-stream PrunedHub-GPT-OSS-20B-28x-Q4_K_M.gguf 512 \
  --prompt "Explain quantum computing" --stream

# OpenAI-compatible HTTP server
moe-stream-server --model PrunedHub-GPT-OSS-20B-28x-Q4_K_M.gguf --port 11434

Methodology

This model was created using GOBA-AI-Labs' expert optimization pipeline:

  • —Calibration-based importance scoring: Expert importance is measured through actual inference behavior on diverse workloads, not just weight magnitude statistics. This produces significantly more accurate importance rankings than static analysis
  • —Layer-adaptive expert allocation: Each MoE layer retains a dynamically determined number of experts based on its measured contribution to model quality, rather than applying a uniform pruning ratio across all layers
  • —Cross-model validation: The pruning methodology has been validated across multiple MoE architectures (GPT-OSS-20B, Qwen3-30B-A3B, Qwen3-Coder-Next-80B) with consistent results

Citation

bibtex
@misc{goba-ai-labs-prunedhub-gptoss-28x,
  title={PrunedHub GPT-OSS-20B-28x: Lossless MoE Expert Pruning},
  author={GOBA-AI-Labs},
  year={2026},
  url={https://huggingface.co/GOBA-AI-Labs/PrunedHub-GPT-OSS-20B-28x}
}

Links