ZeroXClem/Qwen3-4B-Hermes-Axion-Pro

license: apache-2.0 tags:

• —merge
• —mergekit
• —lazymergekit
• —hermes
• —model_stock
• —qwen3-4B datasets:
• —NousResearch/Hermes-3-Dataset language:
• —en base_model:
• —Qwen/Qwen3-4B-Thinking-2507
• —bunnycore/Qwen3-4B-Pro
• —ertghiu256/Qwen3-4b-tcomanr-merge-v2.2
• —ertghiu256/Qwen3-4B-Thinking-2507-Hermes-3 pipelinetag: text-generation libraryname: transformers ---

🧠 ZeroXClem-Qwen3-4B-Hermes-Axion-Pro

Overview

ZeroXClem-Qwen3-4B-Hermes-Axion-Pro is a powerful, safety-conscious, and deeply intelligent merge crafted via Model Stock merging using MergeKit. This 4B-parameter model blends the best of Hermes-3, Axion-Thinking, and Qwen3-Pro, optimized for deep reasoning, safe generation, and dynamic roleplay.

It’s designed to excel in structured problem-solving, multi-turn dialogue, and creative writing, while maintaining safe behavior aligned through red teaming and post-training.

This model is VERY good with reasoning, and hard tasks! Use the default Template(Jinja) setting in LMStudio for best inference.

🔧 Merge Details

• —Model Name: ZeroXClem-Qwen3-4B-Hermes-Axion-Pro
• —Merge Method: model_stock
• —Base Model: `bunnycore/Qwen3-4B-Pro`
• —Dtype: bfloat16
• —Tokenizer Source: `Qwen/Qwen3-4B-Thinking-2507`

YAML Configuration

name: ZeroXClem-Qwen3-4B-Hermes-Axion-Pro
base_model: bunnycore/Qwen3-4B-Pro
dtype: bfloat16
merge_method: model_stock
models:
  - model: ertghiu256/Qwen3-4b-tcomanr-merge-v2.2
  - model: ertghiu256/Qwen3-4B-Thinking-2507-Hermes-3
  - model: Qwen/Qwen3-4B-Thinking-2507
tokenizer_source: Qwen/Qwen3-4B-Thinking-2507

🧬 Models Merged

🧠 `ertghiu256/Qwen3-4B-Thinking-2507-Hermes-3`

Finetuned on the Hermes 3 dataset for instruction alignment and coherent multi-step thinking.

🔒 `AdvRahul/Axion-Thinking-4B`

Safety-tested and enhanced via red teaming protocols. Based on Qwen3-4B-Thinking-2507, with refined behavior for ethical deployment.

🧰 `ertghiu256/Qwen3-4b-tcomanr-merge-v2.2`

Strong logic and instruction-following merge with emphasis on quality output in diverse domains.

💼 `bunnycore/Qwen3-4B-Pro`

A professional-grade Qwen variant tuned for real-world applications like coding, RP, creative writing, and structured tasks.

✨ Features & Highlights

🔹 Deep Thinking & Problem Solving — Inspired by Hermes-3 and Axion, this model handles multi-step logical reasoning and instruction-following with clarity.

🔹 Safe, Aligned Outputs — Red team finetuning and post-training ensure behavior safety and moderation-ready generation.

🔹 Creative Writing & Roleplay — Retains high fluency and character immersion for natural roleplay and storytelling.

🔹 Coding & Engineering Tasks — Competent in code generation, debugging, and technical explanations.

🔹 Efficient & Lightweight — At just 4B parameters, it's easy to deploy locally or in constrained environments.

🎯 Use Cases

• —🤖 Conversational AI
• —✍️ Creative Roleplay & Fiction Writing
• —🧠 Reasoning & Problem-Solving Tasks
• —🧑‍💻 Code Generation & Completion
• —🔐 Safe AI Assistants with Aligned Behavior

🚀 Usage Instructions

For optimal inference, use a higher quant such as [Q6 Here](https://huggingface.co/ZeroXClem/Qwen3-4B-Hermes-Axion-Pro-Q6_K-GGUF).

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ZeroXClem/Qwen3-4B-Hermes-Axion-Pro"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Describe the principles of quantum entanglement in simple terms."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(



    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

For LM Studio Users: When using this model in LM Studio, select the Qwen3-Chat Template from the template dropdown menu. This official template ensures proper prompt formatting for consistent multi-turn conversations and system instruction handling.

⚠️ Alignment & Ethics

• —🔐 Safety Notice: While post-trained with red teaming protocols, this model still outputs raw generations. Always include content moderation for public deployments.
• —🧠 Thinking Mode Support: Fully compatible with enable_thinking=True and /think prompt control.
• —📜 License: Apache 2.0 + governed by the licenses of upstream models.

💌 Feedback & Collaboration

We welcome community feedback, prompts, benchmarks, and merge ideas! Reach out via HF comments or GitHub for collaboration.

ZeroXClem Team | 2025 Buy me a coffee ☕ ---