---
license: apache-2.0
tags:
- merge
- mergekit
- lazymergekit
- qwen3-4b
- qwen3
- zeroxclem
- ZeroXClem / Qwen3-4B-CrystalSonic
datasets:
- reedmayhew/claude-3.7-sonnet-reasoning
- Liontix/claude-sonnet-4-100x
language:
- en
base_model:
- miromind-ai/MiroThinker-4B-DPO-v0.2
- prithivMLmods/Muscae-Qwen3-UI-Code-4B
- FractalAIResearch/Fathom-Search-4B
- Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor
- Qwen/Qwen3-4B-Thinking-2507
pipeline_tag: text-generation
library_name: transformers
---

# 🧠 ZeroXClem-Qwen3-4B-CrystalSonic

## Overview

![CrystalSonicPFP](https://cdn-uploads.huggingface.co/production/uploads/64408cd43e0374802e19f454/dOU_QxMVY1YIlNZexhD-g.png)

**ZeroXClem-Qwen3-4B-CrystalSonic** is an elite 4B-parameter merged model designed for deep reasoning, long-context tool use, structured code generation, and agentic autonomy. Built with **MergeKit's model_stock method**, this crystal-clear fusion draws from powerful contributors like **MiroThinker**, **Muscae-UI**, **Fathom-Search**, and **Claude-distilled reasoning variants**. At its heart lies **Qwen3-4B-Pro**, making this model both versatile and production-ready.

---

## 🔧 Merge Configuration

```yaml
name: ZeroXClem-Qwen3-4B-CrystalSonic
base_model: bunnycore/Qwen3-4B-Pro
dtype: bfloat16
merge_method: model_stock
models:
  - model: miromind-ai/MiroThinker-4B-DPO-v0.2
  - model: prithivMLmods/Muscae-Qwen3-UI-Code-4B
  - model: FractalAIResearch/Fathom-Search-4B
  - model: Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor
  - model: Qwen/Qwen3-4B-Thinking-2507
tokenizer_source: Qwen/Qwen3-4B-Thinking-2507
````

---

## 🧬 Models Merged

### 🧠 [`miromind-ai/MiroThinker-4B-DPO-v0.2`](https://huggingface.co/miromind-ai/MiroThinker-4B-DPO-v0.2)

A cutting-edge agentic model with 64k context, designed for task decomposition, web search, retrieval-augmented reasoning, and long-horizon problem solving. Built on DPO with multilingual capabilities.

### 💻 [`prithivMLmods/Muscae-Qwen3-UI-Code-4B`](https://huggingface.co/prithivMLmods/Muscae-Qwen3-UI-Code-4B)

Fine-tuned for structured code generation in HTML, React, Tailwind, Markdown, and YAML. Supports layout-aware reasoning, component hierarchy, and UI prototyping with structured output.

### 🌍 [`FractalAIResearch/Fathom-Search-4B`](https://huggingface.co/FractalAIResearch/Fathom-Search-4B)

Trained for open-ended, deep information retrieval and autonomous search workflows. Sets new benchmarks in **DeepSearch**, surpassing GPT-4o + Search on reasoning-heavy QA.

### 🎭 [`Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor`](https://huggingface.co/Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor)

Distilled from Claude Sonnet 4/3.7, this model contributes high-fidelity reasoning and conversational engagement to the CrystalSonic blend.

### 🚀 [`Qwen/Qwen3-4B-Thinking-2507`](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507)

Base for long-context thought generation (262k context length). Improved reasoning across logic, math, alignment, tool use, and creativity.

---

## ✨ Features & Highlights

🔹 **Advanced Reasoning & DeepSearch** — From Fathom and MiroThinker: search-aware, long-horizon, tool-augmented thinking.

🔹 **UI & Structured Code Generation** — Muscae-UI brings layout-aware reasoning and polished frontend component synthesis.

🔹 **Safe & Aligned Dialogues** — Claude-style instruction distillation adds emotional nuance and safe defaults.

🔹 **Agentic Capabilities** — Native support for thinking modes, planning, web search, file parsing, and external tool use.

🔹 **Multilingual & Scientific** — Handles technical, scientific, and cross-lingual queries with elegance and depth.

---

## 🎯 Ideal Use Cases

* 🧑‍💻 **Frontend & UI Prototyping**
* 🧠 **Search-Augmented Autonomous Agents**
* 🧬 **Scientific Reasoning & Math**
* 💬 **Conversational AI with Deep Context**
* 📑 **Tool-Augmented Research Assistants**
* 🔍 **Structured Information Synthesis**

---

## 🚀 Quickstart (Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "ZeroXClem/Qwen3-4B-CrystalSonic"

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

prompt = "Explain how quantum computing could impact AI research."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---


📜 **Licenses:**

* Apache 2.0: Credit to MiroThinker, Fathom-Search, Muscae, Qwen3-4B for their amazing models!

---

## 💌 Feedback & Contributions

We welcome your prompts, benchmarks, and merge proposals!

🌐 Hugging Face: [@ZeroXClem](https://huggingface.co/ZeroXClem)
📬 GitHub Issues & PRs: Let’s build smarter agents together.

---

**ZeroXClem Team | 2025**