atlasium-efficient/Qwen3-12B-20pct-Compressed-14B-EN-V1.5

Qwen3-12B — 20% Compressed from Qwen3-14B (English · Chat)

This repository is part of the Efficient and Robust AI System (E-AI) Project by Vincent-Daniel Yun, which publicly releases compressed large language models. This model is a compressed edition of Qwen/Qwen3-14B with 8 of 40 transformer layers removed (32 layers remain, ≈12B parameters), then instruction-tuned for chat so it generates coherent responses at lower memory and latency.

🔗 Project: https://www.worldwidedaniel.com/eai-project 📅 Release date: 2026-07-01 · Version: V1.5

What's new in V1.5

This is the same model as V1 — the weights are unchanged. The only difference is how it writes text.

In V1, the model could repeat itself or drift off when writing longer replies. V1.5 turns on repetition control by default, so you get clean answers with no extra setup. Because of that, it now handles short English chat, not just classification and scoring.

• The repetition problem from V1 is fixed
• Short English chat works with no extra settings
• English only. For questions about specific facts, pair it with retrieval (RAG). It works best with short answers.
• Same benchmark scores as V1, since the weights did not change

Chat examples

A few real replies from this model (English system prompt, default settings):

You: Hello! Can you briefly introduce yourself?

Assistant: Hello! I'm an AI assistant developed by Alibaba Cloud. I'm here to help you with information, assist with tasks, and provide support in a friendly and efficient manner. How can I assist you today?

---

You: Give me three practical tips for staying focused while studying.

Assistant: Sure! Here are three practical tips for staying focused while learning:

1. Set Clear Goals and Use Task Lists: Break your study session into manageable tasks and set specific goals. This helps with clarity and gives you a sense of accomplishment as you complete each one, keeping you motivated.

2. Create a Dedicated Study Space: Set up a clean, quiet environment dedicated to studying. Removing distractions like mobile phones or social media can significantly improve concentration and concentration.

3. Interval Learning with Pomodoro Technique: Work for a fixed period (like 45 minutes) and rest for a short time (about 10 minutes). This pattern helps maintain high levels of focus without overingramming your brain.

These strategies can help you stay focused and productive during study sessions.

---

You: Write a short, friendly message wishing a friend good luck on their exam.

Assistant: Hey buddy,
Just a quick note to say you're absolutely on the right track and I'm all support you're going to do great! Just remember to breathe, think positively, and trust your preparation. Best wishes on your exam — you're a strong and capable individual, and this is just another step in your journey! 🚀
Cheers to success!

P.S. Don’t worry about the results — what matters is how you prepare and how you feel during the exam. You're already halfway already! 💪✨

How to chat

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

name = "daniel-eai/Qwen3-12B-20pct-Compressed-14B-EN-V1.5"
model = AutoModelForCausalLM.from_pretrained(name, trust_remote_code=True, dtype=torch.float16, device_map="cuda")
tok = AutoTokenizer.from_pretrained(name, trust_remote_code=True)

messages = [
    {{"role": "system", "content": "You are a helpful assistant. Always respond in English."}},
    {{"role": "user", "content": "Give me three tips for staying focused while studying."}},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt",
                                 enable_thinking=False).to("cuda")
# repetition control is already set in generation_config.json
output = model.generate(inputs, attention_mask=torch.ones_like(inputs), max_new_tokens=256)
print(tok.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

⚠️ Language support — English only. This model is tuned on English data and is English-focused. Other languages (e.g., Korean, Chinese, Japanese) are not officially supported and may produce degraded or broken output. For open-domain factual questions, use retrieval (RAG) — the compressed model can hallucinate specific facts.

✅ What to use this model for

Our evaluations (full benchmarks vs the dense Qwen3-14B and an off-the-shelf Qwen3-8B) show this compressed model is best used as a fast, low-cost discrimination / classification engine — not as a free-form text generator.

✅ Recommended — matches or beats both the dense 14B and Qwen3-8B here:

• Text classification (sentiment, topic, intent routing)
• Safety & content moderation (toxicity / harmfulness detection — and it is less biased than the dense model)
• Reading comprehension & RAG reading (including multilingual reading)
• Medical / biomedical domain QA (MedQA, MedMCQA, PubMedQA, clinical knowledge, anatomy)
• Natural language inference, paraphrase / duplicate detection, word-sense
• Preference scoring / LLM-as-judge

❌ Not recommended — use the full model instead:

• Open-ended generation / chat, multi-step chain-of-thought reasoning, long-context tasks, math, and broad commonsense reasoning.

About E-AI

Modern AI is powerful but heavy. State-of-the-art models are enormous and their inference is slow — still far from human intuition, and far too slow and unreliable to trust in urgent, high-stakes moments.

Two obstacles stand between today's models and AI we can trust in the field. Individually, each model is too large and too slow to run where it is actually needed. Collectively, when many models or agents work together, a single faulty or adversarial member can quietly derail the whole system. E-AI attacks both — making every model lightweight and fast, and keeping teams of agents reliable even when some of them fail.

I started the E-AI (Efficient-AI) project to build compact yet powerful AI that can assist people in disaster scenarios — responding to dangerous accidents quickly and reliably when every second counts.

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 for chat (distilled from the base model). Only the resulting model is shared here.

Results (measured)

All numbers below were measured by us. PPL is on a 2048-token context (lower is better); downstream tasks and MMLU are 0-shot accuracy via lm-eval-harness (higher is better).

Metric	Qwen3-14B (base)	This model (20%)
PPL · WikiText2 ↓	8.64	19.50
PPL · C4 ↓	13.00	24.41
PPL · PTB ↓	14.79	31.62
ARC-c ↑	0.6024	0.4659
ARC-e ↑	0.8279	0.6987
BoolQ ↑	0.8933	0.8817
COPA ↑	0.9000	0.8400
HellaSwag ↑	0.7881	0.6856
OpenBookQA ↑	0.4620	0.3740
RACE ↑	0.4325	0.4134
RTE ↑	0.7762	0.7184
WinoGrande ↑	0.7317	0.6693
Avg. downstream (9) ↑	0.7127	0.6385
MMLU ↑	0.7729	0.7219

Task-suitability evaluation (when to use this model)

Beyond standard benchmarks, we measured where this 12B compressed model is competitive with — or stronger than — the full Qwen3-14B (dense) and an off-the-shelf Qwen3-8B, on discrimination / classification tasks (full test sets; lm-eval-harness / likelihood scoring, 0-shot).

Real-world discriminative use-cases (vs dense 14B and Qwen3-8B)

Task	This model (12B)	Dense 14B	Qwen3-8B
Topic classification (AG News)	0.866	0.742	pending
LLM-as-judge (RewardBench, pairwise)	0.842	0.874	0.860
Safety — BeaverTails (F1)	0.770	0.786	0.728
Safety — ToxicChat (F1)	0.690	0.667	0.615
Safety — ToxiGen (F1)	0.760	0.755	0.774
SafetyBench (MCQ)	0.857	0.886	0.914

Tasks where this 12B model outperforms Qwen3-8B (full test sets)

Task	This model (12B)	Qwen3-8B	Δ
MultiRC (reading comprehension)	0.572	0.199	+0.372
WiC (word-sense disambiguation)	0.624	0.503	+0.121
ToxiGen (toxicity detection)	0.529	0.455	+0.073
MRPC (paraphrase detection)	0.723	0.657	+0.066
CB (natural language inference)	0.786	0.732	+0.054
XNLI-zh (Chinese NLI)	0.393	0.347	+0.046
Ethics — utilitarianism	0.699	0.672	+0.026
TruthfulQA (mc2)	0.561	0.544	+0.017
BoolQ	0.882	0.866	+0.016
MedQA (USMLE, 4-opt)	0.656	0.641	+0.015
Ethics — virtue	0.902	0.887	+0.015
SST-2 (sentiment)	0.932	0.919	+0.014
Ethics — deontology	0.626	0.612	+0.014
MedMCQA (medical)	0.609	0.596	+0.013
Belebele (Korean reading)	0.843	0.837	+0.007
PubMedQA (biomedical)	0.770	0.764	+0.006
Belebele (English reading)	0.918	0.913	+0.004

MMLU professional/STEM subjects where this model also beats Qwen3-8B: Anatomy +0.059, Electrical Engineering +0.055, Machine Learning +0.045, College Physics +0.029, plus College Medicine, College Biology, and Clinical Knowledge.

Takeaway. This model is well suited as a fast, lightweight discrimination engine — text classification, safety / content moderation, reading comprehension, medical-domain QA, and preference scoring — where it matches or exceeds both the dense 14B and a same-tier 8B model. (For open-ended long-form generation and broad commonsense reasoning, the dense model remains preferable.)

Performance by subject (MMLU-PRO)

MMLU-PRO is a harder, reasoning-focused version of MMLU — 12,032 questions across 14 subjects, answered with step-by-step reasoning. We ran the full test set (no sampling) and compared this model against the dense Qwen3-14B, so you can see where compression costs the least and the most.

Subject	Dense 14B	This model (20%)	Retained
Psychology	0.732	0.665	91%
Economics	0.722	0.635	88%
History	0.583	0.509	87%
Biology	0.806	0.693	86%
Law	0.349	0.287	82%
Health	0.654	0.532	81%
Philosophy	0.549	0.441	80%
Other	0.609	0.463	76%
Computer Science	0.627	0.434	69%
Engineering	0.361	0.198	55%
Physics	0.495	0.223	45%
Math	0.603	0.218	36%
Business	0.598	0.207	35%
Chemistry	0.450	0.125	28%
Overall (official)	0.565	0.367	65%

The model keeps most of its accuracy on knowledge- and reading-heavy subjects (psychology, economics, biology, history, health), and loses the most on multi-step quantitative subjects (chemistry, math, business, physics). In practice, it's a good fit for humanities, social science, and life- and medical-science questions, and a poor fit for heavy calculation.

Model family — pick your size

All sizes in this release (click to open each model). Memory is measured peak inference (fp16 and 4-bit, batch 4 × 2048, single 48 GB GPU).

Model	Layers	Params	MMLU ↑	Avg DS ↑	Mem fp16	Mem 4-bit
Qwen3-14B (base, uncompressed)	40	14.77B	0.773	0.713	33.5 GB	13.9 GB
➡ 20% (this model)	32	12.13B	0.722	0.639	27.9 GB	12.25 GB
Qwen3-11B-25pct-Compressed-14B-EN-V1.5	30	11.47B	0.680	0.598	26.5 GB	11.84 GB
Qwen3-11B-30pct-Compressed-14B-EN-V1.5	28	10.80B	0.626	0.560	25.1 GB	11.44 GB

Compressed + Quantization — GPU memory vs dense

How little GPU memory each option needs relative to the original dense fp16 model (lower is better; combine compression with 4-bit for the largest savings).

Configuration	Peak GPU memory	vs dense fp16
Qwen3-14B dense (fp16)	33.5 GB	100%
20% compressed (fp16) ⬅	27.9 GB	83%
25% compressed (fp16)	26.5 GB	79%
30% compressed (fp16)	25.1 GB	75%
20% compressed + 4-bit ⬅	12.25 GB	37%
25% compressed + 4-bit	11.84 GB	35%
30% compressed + 4-bit	11.44 GB	34%

Efficiency (measured, fp16, batch 4 × 2048, single 48 GB GPU)

	Qwen3-14B (base)	This model (20%)
Layers	40	32
Parameters	14.77B	12.13B
Peak inference memory (fp16)	33.5 GB	27.9 GB (−17%)
Peak inference memory (4-bit)	13.9 GB	12.25 GB (−63% vs dense fp16)
Forward latency (fp16)	2246 ms	1830 ms (−19%)

Quantization

4-bit (and other) quantization can be used with this model — it is a standard Qwen3 architecture, so bitsandbytes 4-bit / 8-bit loading and other PTQ methods apply on top of the compression. Verified: this model loads and generates correctly in 4-bit, with peak inference memory ~12.25 GB (vs 13.9 GB for the dense model in 4-bit, and 33.5 GB for the dense model in fp16).

from transformers import AutoModelForCausalLM, BitsAndBytesConfig
m = AutoModelForCausalLM.from_pretrained(
    "daniel-eai/Qwen3-12B-20pct-Compressed-14B-EN-V1.5", 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(
    "daniel-eai/Qwen3-12B-20pct-Compressed-14B-EN-V1.5", trust_remote_code=True, dtype=torch.float16, device_map="cuda")
tok = AutoTokenizer.from_pretrained("daniel-eai/Qwen3-12B-20pct-Compressed-14B-EN-V1.5", 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="daniel-eai/Qwen3-12B-20pct-Compressed-14B-EN-V1.5", 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

Sincere thanks to Prof. Sai Praneeth Karimireddy (University of Southern California) for his invaluable advice and feedback throughout this work, and to Prof. Sunwoo Lee (Inha University) for his guidance and support. We are also grateful to Alibaba (the Qwen team) for openly releasing the Qwen3-14B base model that made this work possible.