dystrio/Qwen3.5-9B-Sculpt-Experimental

Qwen3.5-9B-Sculpt-Experimental

18% FFN compression with live teacher distillation. Drop-in replacement — no custom kernels, no runtime changes.

Dystrio Sculpt structurally compresses transformer FFN layers, producing dense models that load with standard transformers.

This is the Experimental tier of Qwen3.5-9B.

Use case: Local — maximum compression (1.27x prefill)

Benchmark Results (lm_eval)

This Model vs Baseline

Performance

KV-cache footprint is unchanged — Sculpt only compresses FFN layers, not attention.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "dystrio/Qwen3.5-9B-Sculpt-Experimental",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("dystrio/Qwen3.5-9B-Sculpt-Experimental")

inputs = tokenizer("The future of AI inference is", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

All Sculpt Tiers

Technical Details

• —Method: Structural FFN pruning with importance-aware block selection + live teacher distillation (alpha=0.5)
• —Keep fraction: 0.82 (18% of FFN neurons removed)
• —Repair: 8-stage cosine-LR fine-tuning with best-checkpoint restore
• —Training data: general_v2 mixture (WikiText, OpenHermes 2.5, MMLU, HellaSwag, GSM8K, OpenOrca)
• —Hardware: 1x NVIDIA H200 141GB
• —Output: Standard dense transformer — loads with any HuggingFace-compatible framework

Compatibility

• —HuggingFace Transformers
• —vLLM
• —TGI (Text Generation Inference)
• —llama.cpp / GGUF conversion
• —AWQ / GPTQ quantization
• —Any framework that loads standard safetensors

Citation

@misc{dystrio_sculpt_2026,
  title={Dystrio Sculpt: Structural Compilation for Transformer LLMs},
  author={Dystrio},
  year={2026},
  url={https://huggingface.co/dystrio}
}