Qwen3-12B-20pct-Compressed-14B-EN-V2

Qwen3 — 20% Compressed from Qwen3-14B (English · Chat)

Part of the Efficient and Robust AI System (E-AI) Project by Vincent-Daniel Yun. A compressed edition of Qwen/Qwen3-14B with 8 of 40 transformer layers removed (32 layers remain, ≈12.13B parameters), then instruction-tuned so it runs at lower memory and latency.

📅 Version: V2

What's new in V2

• Full MMLU-Pro reporting — overall score and a per-subject breakdown vs the dense 14B (below).
• Instruction-tuning refresh that improves reasoning-heavy benchmarks over V1/V1.5.
• English only. For open-domain factual questions, pair with retrieval (RAG); best with short answers.

⚠️ Language support — English only. Tuned on English data. Other languages (e.g., Korean, Chinese, Japanese) are not officially supported and may degrade.

Method

The pruning method and the recovery method used to build this model are proprietary, undisclosed methods created by Vincent-Daniel Yun and are not released. The compressed model is then instruction-tuned (distilled from the base model). Only the resulting model is shared.

Results (measured)

PPL on 2048-token context (lower is better); downstream tasks and MMLU are 0-shot accuracy via lm-eval-harness (higher is better).

Performance by subject (MMLU-Pro)

MMLU-Pro is a harder, reasoning-focused MMLU — 12,032 questions across 14 subjects. We ran the full test set (no subsampling) vs the dense Qwen3-14B, via lm-eval-harness: 5-shot, greedy decoding (temperature 0), up to 256 generated tokens (max_gen_toks=256, the harness default), scored by exact match.

Accuracy is best retained on knowledge- and reading-heavy subjects (psychology, economics, biology, health) and lowest on multi-step quantitative subjects (chemistry, math, physics).

Model family — pick your size

Quantization

4-bit / 8-bit quantization works — this is a standard Qwen3 architecture, so bitsandbytes loading and other PTQ methods apply on top of the compression for the largest memory savings.

from transformers import AutoModelForCausalLM, BitsAndBytesConfig
m = AutoModelForCausalLM.from_pretrained(
    "atlasium-efficient/Qwen3-12B-20pct-Compressed-14B-EN-V2", trust_remote_code=True, device_map="cuda",
    quantization_config=BitsAndBytesConfig(load_in_4bit=True))

Usage — Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

m = AutoModelForCausalLM.from_pretrained(
    "atlasium-efficient/Qwen3-12B-20pct-Compressed-14B-EN-V2", trust_remote_code=True, dtype=torch.float16, device_map="cuda")
tok = AutoTokenizer.from_pretrained("atlasium-efficient/Qwen3-12B-20pct-Compressed-14B-EN-V2", trust_remote_code=True)

ids = tok("The capital of France is", return_tensors="pt").to("cuda")
print(tok.decode(m.generate(**ids, max_new_tokens=20)[0]))

trust_remote_code=True is required: the model ships a small custom decoder layer in modeling_qwen3_recovered.py.

Usage — vLLM

vLLM uses its own model implementations, so the custom decoder layer is loaded via a tiny plugin (provided in this repo under vllm_plugin/). Install it once, then serve normally:

pip install ./vllm_plugin   # from a checkout of this repo's vllm_plugin/ folder

from vllm import LLM, SamplingParams
llm = LLM(model="atlasium-efficient/Qwen3-12B-20pct-Compressed-14B-EN-V2", trust_remote_code=True, dtype="float16")
print(llm.generate(["The capital of France is"], SamplingParams(max_tokens=20))[0].outputs[0].text)

Other backends: TGI / SGLang / llama.cpp each use their own model graphs and would need an analogous custom decoder layer; they are not supported out of the box.

License

Apache-2.0, inherited from the base model Qwen/Qwen3-14B.

Acknowledgements

Thanks to Prof. Sai Praneeth Karimireddy (USC) and Prof. Sunwoo Lee (Inha University) for their guidance, and to Alibaba (the Qwen team) for the Qwen3-14B base model.