---
license: apache-2.0
tags:
- merge
- mergekit
- lazymergekit
- qwen3
- ZeroXClem
- hermes
- SkyHighHermes
- Heretic
- Abliterated
- HighReasoning
- Distilled
- Claude
base_model:
- DavidAU/Qwen3-4B-Instruct-2507-Polaris-Alpha-Distill-Heretic-Abliterated
- >-
  DavidAU/Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated
- TeichAI/Qwen3-4B-Thinking-2507-Claude-Haiku-4.5-High-Reasoning-Distill
- TeichAI/Qwen3-4B-Instruct-2507-Claude-Opus-3-Distill
- TeichAI/Qwen3-4B-Thinking-2507-MiMo-V2-Flash-Distill
- TeichAI/Qwen3-4B-Thinking-2507-MiniMax-M2.1-Distill
- Qwen/Qwen3-4B-Thinking-2507
- ZeroXClem/Qwen3-4B-Hermes-Axion-Pro
---

# 🦅 ZeroXClem/Qwen3-4B-Sky-High-Hermes

![SkyHighHermes](https://cdn-uploads.huggingface.co/production/uploads/64408cd43e0374802e19f454/wl5PAvadV2SU-LbMG-z4E.png)


**ZeroXClem/Qwen3-4B-Sky-High-Hermes** is a highly distilled and precision-merged Qwen3-based 4B parameter model optimized for ultra-high reasoning, neutral-aligned autonomy, and elevated conversational depth.

This model is an advanced evolution of `ZeroXClem/Qwen3-4B-Hermes-Axion-Pro`, combining multiple state-of-the-art Heretic Abliterated reasoning experts with Claude 4.5, Gemini 3, Opus 3, and Haiku distillations — all under a finely tuned **262,144 token context window**.

---

## 🧠 Merge Methodology

- **Merge Method**: `model_stock`
- **Base Model**: `ZeroXClem/Qwen3-4B-Hermes-Axion-Pro`
- **Precision**: `bfloat16`
- **Tokenizer Source**: `Qwen/Qwen3-4B-Thinking-2507`
- **Context Length**: `262,144 tokens`

```yaml
name: ZeroXClem/Qwen3-4B-Sky-High-Hermes
base_model: ZeroXClem/Qwen3-4B-Hermes-Axion-Pro
dtype: bfloat16
merge_method: model_stock
models:
  - model: DavidAU/Qwen3-4B-Instruct-2507-Polaris-Alpha-Distill-Heretic-Abliterated
  - model: DavidAU/Qwen3-4B-Thinking-2507-Gemini-3-Pro-Preview-High-Reasoning-Distill-Heretic-Abliterated
  - model: DavidAU/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Heretic-Abliterated
  - model: TeichAI/Qwen3-4B-Thinking-2507-Claude-Haiku-4.5-High-Reasoning-Distill
  - model: TeichAI/Qwen3-4B-Instruct-2507-Claude-Opus-3-Distill
  - model: TeichAI/Qwen3-4B-Thinking-2507-MiMo-V2-Flash-Distill
  - model: TeichAI/Qwen3-4B-Thinking-2507-MiniMax-M2.1-Distill
tokenizer_source: Qwen/Qwen3-4B-Thinking-2507
````

---

## ⚔️ Purpose & Strengths

Sky-High-Hermes is designed to **soar above** traditional instruction-tuned models by harmonizing long context, deep reasoning, and full-spectrum capability.

### ✨ Features

* 🔍 **High-Reasoning Ability** — Claude, Gemini, and MiMo distills support deep multi-hop and abstract reasoning tasks.
* 🧠 **262K Context Window** — Massive memory capacity ideal for longform tasks, coding agents, or narrative tracking.
* 🔓 **Abliterated via Heretic Protocol** — Uncensored, refusal-resistant generation with KL divergence near-zero.
* 💬 **Sublime Dialogue & RP** — High immersion support for fictional characters, story continuations, and dynamic roleplay.
* ⚙️ **Engineering-Grade** — Code writing, debugging, and architectural reasoning.
* 🧬 **Unsloth Optimized** — Fast inference and training via TRL and Unsloth acceleration.

---

## 🛠️ Recommended Use

This is a **Class 1** model suitable for:

* Creative Writing
* Autonomous Agents
* Code Assistance
* Science & Research
* Story & Worldbuilding
* Philosophical Discourse
* NSFW (Uncensored Use)

For optimal inference:

* Set `smoothing_factor = 1.5` (in KoboldCpp, WebUI, or SillyTavern)
* Enable `enable_thinking=True` for structured prompt reasoning

---

## 🚀 Example Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

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

prompt = "Explain the philosophical connection between memory and time."
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=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## 🔐 Licensing & Alignment

* **License**: Apache 2.0
* **Censorship**: Abliterated via Heretic v1.0.1
* **KL Divergence**: ~0.00 – 0.06
* **Refusal Rate**: ≤ 8/100
* **Content Moderation**: Recommended for production use

---

## 💌 Acknowledgements

Gratitude to:

* **DavidAU** – For crafting and liberating high-reasoning uncensored blends
* **TeichAI** – For high-resolution Claude/Gemini distillations
* **Unsloth Team** – For inference-accelerated Qwen3 tuning frameworks
* **Mergkit Acree Team** - For providing such a fantastic framework to merge these LLMS
* **Alibaba Qwen Team** - For providing such amazing open source base models

---

🕊️ *Soar higher. Think freer. Respond truer.*
Built with love by **ZeroXClem | 2026**

```