🧠 ZeroXClem-Qwen3-4B-CrystalSonic

Overview

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

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`

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`

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`

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`

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`

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)

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))