crownelius/Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5

🪶 CROW-4B

Flagship Intelligence. Featherweight Footprint. Meticulously distilled from Claude Opus 4.6 into a highly efficient Qwen 3.5 architecture.

---

Architecture: Qwen 3.5 | Parameters: 4 Billion | Teacher Model: Claude Opus 4.6 | Type: Distilled LLM

---

🌟 Model Highlights

• Distilled Excellence: Captures the deep reasoning, nuanced formatting, and instruction-following capabilities of Claude Opus 4.6.
• Ultra Compact: At just 4B parameters, Crow-4B runs on virtually any consumer GPU or CPU — including laptops and edge devices — without sacrificing contextual depth.
• Qwen 3.5 Backbone: Inherits robust multilingual support, a massive context window, and structural stability.
• Broadest Training Mix: Trained on 15 datasets spanning reasoning, creative writing, agentic coding, security research, and roleplay — over 25,000 carefully curated examples.

---

Available Model Files:

• Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5.BF16.gguf
• Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5.Q8_0.gguf
• Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5.Q5_K_M.gguf
• Crow-4B-Opus-4.6-Distill-Heretic_Qwen3.5.Q4_K_M.gguf

---

Training Details

Base Model	tvall43/Qwen3.5-4B-heretic
Training Framework	Unsloth + TRL SFTTrainer
LoRA Rank	r=32, α=32
Learning Rate	2e-4 (cosine schedule)
Epochs	1
Max Sequence Length	2048 (model max)
Hardware	NVIDIA A100 40GB

Datasets (15 total, ~25,500 examples)

#	Dataset	Category
1	crownelius/Opus4.6-No-Reasoning-260x	Reasoning
2	crownelius/Opus-4.6-Reasoning-2100x-formatted	Reasoning
3	crownelius/Creative_Writing_Multiturn_Enhanced	Creative Writing
4	peteromallet/dataclaw-peteromallet	Agentic
5	TeichAI/Pony-Alpha-15k	Creative / Roleplay
6	crownelius/Pony-Alpha-3000x-formatted	Creative / Roleplay
7	HuggingFaceH4/llava-instruct-mix-vsft	Instruction Following
8	reedmayhew/LittleMonster	Uncensored
9	tandevllc/hacking-tricks	Security Research
10	m-a-p/Code-Feedback	Coding
11	WithinUsAI/Python_GOD_Coder_50k	Coding
12	peteromallet/my-personal-codex-data	Agentic / Coding
13	misterkerns/my-personal-claude-code-data	Agentic / Coding
14	Roman1111111/gemini-3-pro-10000x-hard-high-reasoning	Reasoning
15	reedmayhew/gemini-3.1-pro-2048-reasoning-1100x	Coding + Reasoning

This was trained 2x faster with Unsloth