Qwen3-4B-MiniMight

🧠 ZeroXClem/Qwen3-4B-MiniMight

"Small in scale. Mighty in mind."
A beautifully blended 4B model with 262k context length, fusing deep reasoning, code, safety, and creativity — distilled through MergeKit’s model_stock magic.

---

🔧 Merge Configuration

name: ZeroXClem/Qwen3-4B-MiniMight
base_model: Qwen/Qwen3-4B-Thinking-2507
dtype: bfloat16
merge_method: model_stock
models:
  - model: TeichAI/Qwen3-4B-Thinking-2507-Claude-4.5-Opus-High-Reasoning-Distill
  - model: bunnycore/Qwen3-4B-Mini-Merge
  - model: bunnycore/Qwen3-4B-MegaMerge
  - model: bunnycore/Qwen3-4B-Max-Merge
  - model: ertghiu256/qwen3-4b-claude-sonnet-x-gemini-reasoning
  - model: ZeroXClem/Qwen3-4B-Hermes-Axion-Pro
tokenizer_source: Qwen/Qwen3-4B-Thinking-2507

---

🌟 Highlights

• 🧠 Reasoning-Centric Intelligence — Optimized for multi-step thought, explanation, STEM logic, and symbolic problem solving.
• 🪐 262,144 Token Context Window — Massive memory span for long documents, complex chains, or multi-turn workflows.
• 🧬 Deep Merge of Code, RP, and Claude/Gemini Logic — Captures both fluency and fidelity in multi-domain generation.
• 🔐 Safe Alignment from Hermes & Axion — Includes red-teamed safety-conscious merges to guide outputs.
• ✍️ Elegant Dialogue & Creative Thought — Engages in immersive roleplay, structured writing, and storytelling.
• ⚡ Efficient 4B Inference — Small enough for local GPUs, quantized formats, and mobile deployments.

---

🧪 Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("ZeroXClem/Qwen3-4B-MiniMight", device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("ZeroXClem/Qwen3-4B-MiniMight")

messages = [{"role": "user", "content": "Explain quantum tunneling with an example."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=384)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🧠 Use enable_thinking=True or /think with compatible frontends for enhanced reasoning control.

---

🔁 Models Merged

---

💼 Ideal For

• 🔍 Chain-of-thought reasoning & symbolic logic
• 🧪 Long-document summarization & analysis
• 💬 Roleplay & storytelling with high memory
• 📚 Tutoring & educational AI (math, code, science)
• ⚖️ Safety-aligned assistants for deployment
• 👨‍💻 Code generation, refactoring, and walkthroughs

---

🚫 Limitations

• Long responses may truncate if max_new_tokens is too low
• Mixed training sources may result in style variance across domains
• Dialogue tone defaults to helpful/instructive rather than emotional/creative

---

📜 License

Apache 2.0 Please verify license alignment when adapting for commercial or public-facing use.

---

💌 Credits

Thank you to bunnycore, Qwen Team, ertghiu256, TeichAI for the wonderful fine tunes and base models that made this merge possible! Built by ZeroXClem Team — merging distillations, hugging intellect and ideas into tangible weights. MergeKit mastery allows full distilled insight in one mighty little package.

Mini in name. Infinite in mind. 🛸